← arXiv Toy PoCs · Landscape dashboard
Why these papers matter
A plain-language reading guide to recent arXiv papers, so you can decide which are worth reading in depth. Each entry says what the paper claims, why it would matter if true, who should explore it further, and what to doubt. These are screening judgments, mostly from abstracts, and most papers here have not been tested by us; tested ones link to their findings. Written by Claude (an AI model), directed by vishwasaidev-dev, and Claude can make mistakes: open an issue to correct one. Not affiliated with the papers' authors. Updated 2026-11-17.
Read closely (227)
Pretty Linear Liars: Unregularized Probes Favor Intervention-Unfaithful Features
What the paper claims. A linear probe fitted without regularisation latches onto low-variance directions that merely correlate with the target, so the feature it finds is decodable but the model does not actually use it; ridge regularisation reduces this.
Why it matters if true. If true, probing results that did not use regularisation are weaker evidence of what a model uses than they look, and a cheap fix (ridge) improves causal faithfulness.
Explore deeper if. you read or write probing papers, because the claim is testable in a small toy network by comparing probe direction against an intervention on the same direction.
What to doubt. Evidence is theory plus a toy setting; the effect size on real LLM probes and the choice of ridge strength are open.
Screened from the abstract only, not tested. A reading guide, not a verification.
Attention Sinks as Spectral Spikes: A Mechanism Analysis of Gated Attention
What the paper claims. Attention sinks can be described as low-rank spikes in the token covariance, and output gating works by shrinking those spikes in the forward pass and the gradients routed through them.
Why it matters if true. If true, it gives one account linking sinks, massive activations and why gated attention trains better, which could guide architecture choices.
Explore deeper if. you train small transformers and can check whether gating changes the top singular values of token covariance.
What to doubt. A mechanism analysis from one framing; the abstract does not say whether it predicts anything new.
Screened from the abstract only, not tested. A reading guide, not a verification.
Sage S2.1 Flash Technical Report
What the paper claims. A 320B science-focused mixture-of-experts model with encoders for molecules and biological sequences.
Why it matters if true. If true, one model reasons across scientific modalities.
Explore deeper if. you work on scientific foundation models at scale.
What to doubt. Technical report; the model name and version in the abstract are inconsistent.
Screened from the abstract only, not tested. A reading guide, not a verification.
K2-Horizon: A Frontier Fleet of Open Source Language Models
What the paper claims. A family of six fully open language models from 0.9B to 375B mixture-of-experts for reasoning and agentic work.
Why it matters if true. If true, open weights across a wide size range.
Explore deeper if. you want open models of many sizes.
What to doubt. Model report with the authors' benchmark comparisons.
Screened from the abstract only, not tested. A reading guide, not a verification.
Where Should LLM Agents Adapt at Test Time? In Context, in Weights, or Both
What the paper claims. Compares letting an agent keep experience in context, update its weights (GRPO or self-distillation), or both, on repeated tasks in the same environment.
Why it matters if true. If the comparison is fair, it tells builders of repeated-use agents which adaptation channel to invest in.
Explore deeper if. you build agents for one codebase or database and must pick between longer memory and fine-tuning.
What to doubt. Agent benchmark results depend heavily on the task streams, base model and compute; likely not reproducible at toy scale.
Screened from the abstract only, not tested. A reading guide, not a verification.
AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data
What the paper claims. Generates portfolio allocation actions end to end from raw market data instead of separate forecasting and construction steps.
Why it matters if true. If true, avoids the mismatch between forecasting and portfolio goals.
Explore deeper if. you study quantitative finance.
What to doubt. Backtest claims in finance are prone to overfitting; not investment advice.
Screened from the abstract only, not tested. A reading guide, not a verification.
From Facts to Personas: Interpretable Role Unlearning in LLMs via Mixture-of-Experts
What the paper claims. Maps different role-play personas to separate expert subspaces in a mixture-of-experts so a target persona's style and knowledge can be forgotten while other personas stay intact.
Why it matters if true. If true, persona removal for privacy or safety would be targeted rather than blunt, and relates to our own persona-forget tests.
Explore deeper if. you work on unlearning or persona control, because it can be compared with simpler adapter-removal baselines.
What to doubt. Forgetting often only hides behaviour; the abstract does not say whether extraction attacks were tried.
Screened from the abstract only, not tested. A reading guide, not a verification.
Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering
What the paper claims. Applying an SAE steering signal on every token hurts fluency; choosing which tokens to steer under a fixed budget, with gates that select or adapt over the generation, keeps control while preserving quality.
Why it matters if true. If true, steering becomes usable without degrading text, by touching fewer tokens.
Explore deeper if. you use activation steering and can test steering a subset of tokens on a small model.
What to doubt. Depends on SAE quality and the behaviours chosen; fixed token budgets may differ from real deployment.
Screened from the abstract only, not tested. A reading guide, not a verification.
Direction or Magnitude? Learning Scalars over Task-Calibrated Directions
What the paper claims. Fine-tuning only a handful of scalar coefficients on top of a fixed basis taken from task activations and random output vectors is said to match LoRA at far fewer trainable parameters.
Why it matters if true. If true, adapters could be tiny enough for per-user personalisation on small devices, with lower memory than LoRA.
Explore deeper if. you care about extremely small adapters or want to test how much of fine-tuning is just rescaling, which a small model can check.
What to doubt. Comparable-to-LoRA claims depend on which benchmarks and budgets were chosen; random output dictionaries working this well may say the tasks are easy.
Screened from the abstract only, not tested. A reading guide, not a verification.
Folding Language Models into Themselves
What the paper claims. Builds a small student by folding groups of teacher layers into one shared block run several times, with low-rank corrections, then recovers capability by distillation.
Why it matters if true. If true, small models could be made from large ones with far less data than training from scratch.
Explore deeper if. you compress models and can compare against pruning or layer-sharing baselines at small scale.
What to doubt. Results reported on Qwen-scale models; the compute and data recipe are not toy-testable.
Screened from the abstract only, not tested. A reading guide, not a verification.
Envisioning Agentic Robotics: A New Era of Robotic Systems
What the paper claims. A position paper arguing that agentic robotics, a foundation model in a closed loop with tools, now matches task-specific policies.
Why it matters if true. If the premise holds, robotics research shifts from per-task policies to agent harnesses.
Explore deeper if. you work in robotics strategy.
What to doubt. A vision paper; the premise about a named model release is taken from the paper itself and unverified here.
Screened from the abstract only, not tested. A reading guide, not a verification.
Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent
What the paper claims. A generalist vision-language-action agent for long-horizon game play, trained on tens of thousands of hours of keyboard and mouse data.
Why it matters if true. If true, games give scalable training for general interactive agents.
Explore deeper if. you build computer-use or game agents.
What to doubt. Large-scale report; benchmark choices are the authors'.
Screened from the abstract only, not tested. A reading guide, not a verification.
Data-Efficient LLM Safety Alignment Guided by Interpretable Features
What the paper claims. A 2.4K-sample safety-alignment set chosen by sparse-autoencoder features outperforms much larger sets on safety and over-refusal benchmarks.
Why it matters if true. If true, safety alignment can be cheaper and cause less over-refusal.
Explore deeper if. you build safety fine-tuning data.
What to doubt. Benchmark wins on nine benchmarks chosen by the authors; selection by SAE features may overlap with benchmark content.
Screened from the abstract only, not tested. A reading guide, not a verification.
When and How SignSGD Outperforms SGD: A Functional Scaling Law Perspective
What the paper claims. Adding gradient noise makes the sign operator behave like a preconditioner, so SignSGD learns the signal faster than SGD but forgets noise more slowly; which wins depends on how hard the task is.
Why it matters if true. If true, the choice between sign-based optimizers (Lion, signSGD) and SGD becomes predictable from the task's spectrum and noise level rather than trial and error, which affects anyone picking an optimizer for a new problem.
Explore deeper if. you tune optimizers or train small models, because the theory is stated for power-law kernel regression and can be checked on a toy linear problem on CPU in minutes.
What to doubt. The result is proved for diagonal-feature kernel regression only; whether the regime boundaries transfer to deep nets with Adam-style second moments is unshown.
Screened from the abstract only, not tested. A reading guide, not a verification.
Maximizing Capacity Utilization of LoRA Adapters for Mitigating the Learning-Forgetting Trade-off
What the paper claims. Standard LoRA updates use far fewer directions than their configured rank allows; adding a regulariser that raises the update's effective rank reduces the learning-versus-forgetting trade-off.
Why it matters if true. If true, a one-line loss term would let LoRA match full fine-tuning more often and forget less, which matters for continual adaptation.
Explore deeper if. you fine-tune with LoRA or study forgetting, because effective rank is cheap to measure on a small model you train yourself.
What to doubt. Gains are reported by the authors on their chosen tasks; no mention in the abstract of how sensitive the result is to the regulariser weight.
Screened from the abstract only, not tested. A reading guide, not a verification.
When can an interpretability claim be trusted? Stability and certification of feature dimensionality, ridge leverage and the feature recovery rate
What the paper claims. Two common interpretability measurements, feature dimensionality and feature recovery rate, can flip with small weight or threshold changes; ridge-regularised versions come with provable stability bounds.
Why it matters if true. If true, safety claims such as 'the dictionary recovered most features' need stability checks before they can be trusted.
Explore deeper if. you report sparse-dictionary recovery numbers, because the instability can be checked on a toy model with known ground truth.
What to doubt. Studied in toy models and synthetic dictionaries; real LLM dictionaries may behave differently.
Screened from the abstract only, not tested. A reading guide, not a verification.
An Operationalization of Opaque Serial Depth
What the paper claims. Proposes a measurable proxy for how much hidden serial reasoning a model can do without passing through human-readable steps, by defining which computation nodes count as interpretable bottlenecks.
Why it matters if true. If it holds, safety teams could track whether new architectures erode chain-of-thought monitorability before deploying them.
Explore deeper if. you work on chain-of-thought monitoring or oversight evaluations.
What to doubt. It is a definition and operationalisation; whether the number predicts real monitoring failures is not shown in the abstract.
Screened from the abstract only, not tested. A reading guide, not a verification.
VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender
What the paper claims. A 242-task benchmark for coding agents that rebuild 4D scenes in Blender and render proxy videos.
Why it matters if true. If true, tests whether coding agents can plan multi-shot scenes for video models.
Explore deeper if. you build Blender or scene-generation agents.
What to doubt. A benchmark measuring the authors' chosen tasks.
Screened from the abstract only, not tested. A reading guide, not a verification.
SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models
What the paper claims. A benchmark mapping the capability spectrum and decision reliability of fast one-pass decision models.
Why it matters if true. If true, shows where fast models fail beyond aggregate scores.
Explore deeper if. you evaluate one-pass models.
What to doubt. A benchmark; results show gaps, not a method.
Screened from the abstract only, not tested. A reading guide, not a verification.
Mechanistic Interpretability Needs a Logic of Inference
What the paper claims. Mechanistic interpretability needs a stricter logic of inference, so the paper proposes five principles borrowed from neuroscience for going from observations to mechanism claims.
Why it matters if true. If adopted, interpretability findings would be easier to compare and less likely to be stories.
Explore deeper if. you write or review interpretability results. Check first whether the principles would have caught known failed claims.
What to doubt. Conceptual, with no experiments in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Few-Step Streaming Video Generation via Condition Shifting
What the paper claims. Few-step streaming video generation without a separate fake-score network, by shifting the condition to estimate the generated distribution within the generator.
Why it matters if true. If true, distillation for video needs less memory and compute.
Explore deeper if. you distil video generators.
What to doubt. Needs large GPUs to reproduce.
Screened from the abstract only, not tested. A reading guide, not a verification.
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
What the paper claims. Agent memory can hurt by storing lessons inferred from outcome-only labels; linking each lesson to its evidence reduces cases where memory makes an agent fail on tasks it would otherwise solve.
Why it matters if true. If true, anyone building self-improving agents gets a cheap safeguard against memory-induced regressions.
Explore deeper if. you build agents that keep lessons across tasks.
What to doubt. Agent benchmarks are noisy and the evidence-linking details decide whether the fix is general.
Screened from the abstract only, not tested. A reading guide, not a verification.
A Geometric Invariant of Attention Predicts Head Specialization from Weights Alone
What the paper claims. A single number computed from an attention head's weights alone (no data, no forward pass) is said to predict which heads become induction heads, in models that use rotary position encoding.
Why it matters if true. If true, you could find the in-context-copying heads of a model by reading its weights, and decide which heads need long-context scaling treatment, without running it. It would make circuit-finding much cheaper.
Explore deeper if. you study induction heads or train small transformers yourself, because the claim can be checked on a tiny RoPE model trained from scratch. Check first that the correlation (-0.55 to -0.74) holds with a head-level control, not just across pooled heads.
What to doubt. Only moderate Spearman correlations, four bias-free RoPE models, and the YaRN application is called a preliminary proof of concept. Correlation across heads may be driven by layer position.
Screened from the abstract, not tested. A reading guide, not a verification.
Learning Heterogeneous Tactile Representation for General Dexterous Manipulation
What the paper claims. Learns a tactile representation that works across different tactile sensor types, so models can use heterogeneous tactile data.
Why it matters if true. If true, dexterous manipulation models stop being tied to one sensor.
Explore deeper if. you work on tactile or dexterous manipulation.
What to doubt. Needs hardware and sensor data to verify.
Screened from the abstract only, not tested. A reading guide, not a verification.
STABLE: A Continual Learning Optimizer with Adaptive Drift Control
What the paper claims. STABLE, an optimizer that holds parameter drift to a budget with a dual variable, lets plain PEFT match specialized continual-learning methods.
Why it matters if true. If true, forgetting could be handled in the optimizer instead of with replay or new architectures.
Explore deeper if. you do continual learning and want one drop-in change. Check first at which drift budgets plasticity is lost.
What to doubt. The theory is for distance-based classifiers, and experiments are on CLIP-style streams.
Screened from the abstract, not tested. A reading guide, not a verification.
Can small transformers learn the language of Coxeter groups?
What the paper claims. A small transformer trained on the language of one Coxeter group (affine A2, the symmetries of the triangle tiling) showed attention patterns that led mathematicians to formulate and prove a new word-theoretic conjecture.
Why it matters if true. It is a rare case where reading a model's attention produced a checked mathematical result, not just a plausible story. If it holds up, it gives a recipe for model-assisted conjecture finding.
Explore deeper if. you like small models on exact, fully checkable domains. The task is cheap to set up and the group is small enough to verify any claim by brute force. Check first how much of the conjecture came from the model and how much from the authors' own insight.
What to doubt. The abstract is brief, with no numbers on model accuracy and no account of how many patterns were examined before one gave a conjecture.
Screened from the abstract, not tested. A reading guide, not a verification.
The Surprising Effectiveness of Deleting Weights in LLM Post-Training
What the paper claims. Setting under 0.05% of a pretrained LLM's weights to zero (a learned binary mask on the gate projection) matches full fine-tuning on reasoning RL, SFT and distillation, while forgetting less.
Why it matters if true. If true, new skills can be stored as a tiny mask hundreds to thousands of times smaller than a checkpoint, and switched on or off, which helps continual learning and model storage.
Explore deeper if. you work on fine-tuning, forgetting or weight localization. It also bears directly on our PEFT-lab questions about where a skill lives in the weights. Check first how it compares with LoRA at equal parameter count.
What to doubt. Results are at a scale we cannot reproduce on CPU, the abstract gives no per-task numbers, and a mask trained against one benchmark may be tuned to it.
Screened from the abstract, not tested. A reading guide, not a verification.
ConwayNet: Interpretability, Activation Coherence, and Noise Resilience via a Game-of-Life-Inspired Lateral Inhibition in Artificial Neural Networks
What the paper claims. A fixed lateral-inhibition rule modeled on Conway's Game of Life, added to ordinary networks without being trained, makes activations spatially coherent and more noise-resistant.
Why it matters if true. If true, a hand-set, non-learned rule can shape what a trained network learns, which echoes our written-weights work.
Explore deeper if. you care about imposed structure in networks. Check first whether accuracy under noise improves against a plain lateral-inhibition baseline.
What to doubt. Claims centre on interpretability-like properties that are hard to measure, and the abstract gives no accuracy numbers.
Screened from the abstract, not tested. A reading guide, not a verification.
Scaling Laws for SAE Training Data
What the paper claims. A sparse autoencoder does not need fresh activations throughout training: past a buffer-size plateau, replaying a smaller buffer is nearly as good, cutting storage 8 to 64 times.
Why it matters if true. If true, SAE training gets far cheaper in storage, so more groups can train their own.
Explore deeper if. you train SAEs on limited disk. Check first that interpretability scores, not just reconstruction, hold at the smallest buffer.
What to doubt. Tested on Llama and Qwen models of unspecified size ranges and on math and code only.
Screened from the abstract, not tested. A reading guide, not a verification.
Encoding and Propagation of Grammatical Number in Large Language Models
What the paper claims. SAE features in Pythia-70M can be traced for plural nouns, plural subjects and plural verbs, and removing them changes the verb the model produces.
Why it matters if true. If true, a small model's grammar is concretely readable, which is a clean testbed for causal feature work.
Explore deeper if. you want a small, CPU-friendly interpretability case study. Check first how many other number-agreement prompts the same features explain.
What to doubt. One phenomenon in one 70M model, and feature interventions can change many things besides number.
Screened from the abstract, not tested. A reading guide, not a verification.
Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement Learning
What the paper claims. In LoRA training, RL concentrates the singular spectrum of the update and rotates its subspace faster than SFT, and the cause is group-mean baseline subtraction.
Why it matters if true. If true, SFT and RL should not share LoRA settings, and the paper gives stage-specific fixes that flatten the spectrum.
Explore deeper if. you fine-tune with LoRA under both SFT and RL. The quadratic multi-index toy model in the paper can be run cheaply. Check first whether the effect appears without math-reasoning data.
What to doubt. The mechanism is argued from a toy model, and gains from the proposed fixes are not summarized in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
What the paper claims. Decodes visual plans such as points, boxes and trajectories in parallel using their known structure, instead of one coordinate token at a time.
Why it matters if true. If true, physical-AI models plan faster without losing accuracy.
Explore deeper if. you build vision-language planners for robots or driving.
What to doubt. Speed-up claims depend on the structure assumed; compare against strong autoregressive baselines.
Screened from the abstract only, not tested. A reading guide, not a verification.
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
What the paper claims. Continual pre-training of a vision-language-action model on broad robot experience to add 3D spatial understanding as a transferable capability.
Why it matters if true. If true, robot models get spatial skill before task-specific tuning.
Explore deeper if. you build VLA models.
What to doubt. Needs robot data and compute to verify.
Screened from the abstract only, not tested. A reading guide, not a verification.
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
What the paper claims. Feed-forward 3D Gaussian splatting predicts scale targets in a radial-distance space that mismatches neighbour spacing; normalising spacing fixes scale shifts across scenes.
Why it matters if true. If true, feed-forward 3DGS generalises better across scene scales.
Explore deeper if. you work on feed-forward 3D reconstruction.
What to doubt. Needs 3D datasets and GPUs to verify.
Screened from the abstract only, not tested. A reading guide, not a verification.
Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
What the paper claims. A streaming speech recogniser that never edits earlier output while keeping accuracy and latency, so downstream translators and voice agents can act on each token safely.
Why it matters if true. If true, voice agents avoid wasted work or wrong actions caused by revised transcripts.
Explore deeper if. you build voice agents or live translation on streaming ASR.
What to doubt. Never-revising output usually costs accuracy at some latency; the abstract's claim of both needs checking on your data.
Screened from the abstract only, not tested. A reading guide, not a verification.
FrontierPhysics: Evaluating Agents for End-to-End Physics Research
What the paper claims. A benchmark of 56 end-to-end research tasks written by 42 physicists that checks research quality as well as correctness.
Why it matters if true. If true, tests agents on real physics research instead of fixed questions.
Explore deeper if. you build or evaluate science agents.
What to doubt. A benchmark; task authorship and grading choices decide its value.
Screened from the abstract only, not tested. A reading guide, not a verification.
From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence
What the paper claims. A perspective that AI should sit inside human cognition as a symbiont that helps people remember, notice and decide, instead of acting alone.
Why it matters if true. A framing for human-AI interaction research.
Explore deeper if. you design assistive AI.
What to doubt. A position paper; no tested claim.
Screened from the abstract only, not tested. A reading guide, not a verification.
Behind Video Generation: Observations and Insights on Multimodal Understanding from Kling
What the paper claims. A report arguing video generation quality is the projection of multimodal understanding through data, conditioning, alignment and evaluation.
Why it matters if true. A practitioner perspective on where video generation is bottlenecked.
Explore deeper if. you work on video generation.
What to doubt. An observations report; few testable claims.
Screened from the abstract only, not tested. A reading guide, not a verification.
Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
What the paper claims. A study of how to route among many models, tools and policies in one general-purpose AI system.
Why it matters if true. If true, orchestration of specialists becomes a measurable design problem.
Explore deeper if. you build routing or mixture-of-agents systems.
What to doubt. Depends on candidate pool and evaluation set.
Screened from the abstract only, not tested. A reading guide, not a verification.
Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training
What the paper claims. Rescues failures of a base agent using sibling harness variants, then transfers those capabilities into one policy.
Why it matters if true. If true, agent training benefits from multiple harnesses without needing a universally best one.
Explore deeper if. you train agents against changing harnesses.
What to doubt. Agent-training results are hard to compare and expensive to reproduce.
Screened from the abstract only, not tested. A reading guide, not a verification.
Think Less, Locally: Local Preference Learning for Reasoning Compression
What the paper claims. Reasoning models over-deliberate after they already have enough information; training on the local decision of whether to continue reasoning compresses chains of thought.
Why it matters if true. If true, inference cost for reasoning models falls without losing accuracy on easy cases.
Explore deeper if. you serve or train reasoning models and pay for long chains of thought.
What to doubt. Reasoning compression results depend on the benchmark mix and the accuracy-length trade-off chosen.
Screened from the abstract only, not tested. A reading guide, not a verification.
CoworkOS:Scaling Agent Coworkers for Real-world Automation
What the paper claims. A harness for long-horizon cowork agents that separates the model's strategy from execution management.
Why it matters if true. If true, long-running office-style automation becomes more reliable.
Explore deeper if. you build long-horizon agents.
What to doubt. System paper with the authors' own benchmark.
Screened from the abstract only, not tested. A reading guide, not a verification.
MoviEdit: Long-form Multi-Shot Video Editing Made Easy
What the paper claims. An end-to-end framework for long multi-shot video editing that keeps the edit coherent across scene and subject changes.
Why it matters if true. If true, editing stays consistent across shots.
Explore deeper if. you work on video editing models.
What to doubt. Needs large GPUs and human review.
Screened from the abstract only, not tested. A reading guide, not a verification.
P$^2$-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors
What the paper claims. Jointly optimising the mixing coefficients and the student weights makes multi-teacher on-policy distillation faster by starting from a task-arithmetic combination.
Why it matters if true. If true, merging several expert models into one becomes cheaper than distilling from scratch.
Explore deeper if. you merge or distill several fine-tuned models.
What to doubt. Merging gains are model- and task-specific; no ablation visible in the abstract.
Screened from the abstract only, not tested. A reading guide, not a verification.
Advances in Agentic Computer Graphics
What the paper claims. A survey of how LLM agents generate and refine graphics code, arguing graphics is becoming agent-built.
Why it matters if true. A map of the field for people entering agentic graphics.
Explore deeper if. you work on graphics or code-generating agents and want a literature overview.
What to doubt. A survey; contains no new tested claim.
Screened from the abstract only, not tested. A reading guide, not a verification.
Wave Forcing: Towards Speed-of-Light Streaming Video Generation
What the paper claims. Real-time long-video generation with multi-step denoising through a co-designed generation, training and execution scheme that avoids double forward passes and camera drift.
Why it matters if true. If true, streaming video generation gets faster and holds quality over long rollouts.
Explore deeper if. you work on real-time video generation and have GPUs.
What to doubt. Quality and speed claims on the authors' benchmarks; not toy-testable.
Screened from the abstract only, not tested. A reading guide, not a verification.
Words Caught in the Heart: Simulating Person-Specific Expression for Self-Opaque Clients in Psychotherapy
What the paper claims. A simulator of psychotherapy clients whose expression is person-specific and not over-clear about their own distress.
Why it matters if true. If true, therapist training and evaluation can use more realistic simulated clients.
Explore deeper if. you work on mental-health AI, with ethical review.
What to doubt. Realism of a simulated patient is hard to validate without clinicians.
Screened from the abstract only, not tested. A reading guide, not a verification.
KV-Cache Patching: Tracing How and Where Fine-Tuned Behaviors Arise
What the paper claims. Patching the KV cache from a fine-tuned model into the base model transfers some behaviors (answering in German, emergent misalignment) and the effect concentrates on a few pivotal positions such as system-prompt and template tokens.
Why it matters if true. If true, some fine-tuned behavior lives in how earlier tokens are processed, not only in generation, which changes where to look when auditing a fine-tune.
Explore deeper if. you audit LoRA fine-tunes or study where behavior is stored. Check first that the effect is not only the chat-template tokens.
What to doubt. Applies only to some behaviors, and the abstract does not say how many models were tested.
Screened from the abstract, not tested. A reading guide, not a verification.
Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
What the paper claims. In a language model trained with RL to play Tic-Tac-Toe, accuracy rises but sparse autoencoder features show the model mostly relies on information already in the prompt, with board symmetries and forks weakly represented.
Why it matters if true. If true, RL gains on reasoning benchmarks may overstate what the model has learned, because the model learns shortcuts that read off the prompt.
Explore deeper if. you evaluate reasoning gains or use SAEs to judge what a model understands. A fully enumerable game makes the claim checkable in a way real benchmarks are not. Check first whether a linear probe agrees with the SAE reading.
What to doubt. Absence of a feature in an SAE is weak evidence of absence in the model, and a 3x3 game may be too simple to generalize from.
Screened from the abstract, not tested. A reading guide, not a verification.
Proving and Improving the Platonic Representation Hypothesis with JEPAs
What the paper claims. Proves that JEPA models trained separately on different modalities end up with the same representation as they scale, turning the Platonic Representation Hypothesis into a theorem, and uses it to suggest an improvement.
Why it matters if true. If true, convergence of representations across modalities becomes a guarantee for this model family, which would justify cross-modal transfer and alignment methods.
Explore deeper if. you study representation similarity or JEPA-style training, because the theorem's conditions can be tested on small paired-modality toy data.
What to doubt. A theorem holds under its stated assumptions (architecture, data, scale limit); the abstract does not show how close real models come to those assumptions.
Screened from the abstract only, not tested. A reading guide, not a verification.
Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement
What the paper claims. Studies the intersection of agentic video generation and video world models through an executable agent harness.
Why it matters if true. A framing for combining planning, tools and video models.
Explore deeper if. you work on video agents.
What to doubt. Position and survey style; few tested claims.
Screened from the abstract only, not tested. A reading guide, not a verification.
Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
What the paper claims. Argues progress toward AGI may come from a light System-1 predictor that screens candidate actions so System-2 only executes worthwhile ones.
Why it matters if true. If true, cheaper agents by filtering before acting.
Explore deeper if. you build agent systems with a screening step.
What to doubt. Broad AGI framing with the authors' own evidence.
Screened from the abstract only, not tested. A reading guide, not a verification.
A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement
What the paper claims. A zero-data multimodal self-improvement method that decides when to refresh training tasks instead of only what to generate, avoiding majority-vote targets.
Why it matters if true. If true, self-improvement loops avoid stalling or self-confirming.
Explore deeper if. you study self-improving models.
What to doubt. Self-improvement gains are hard to separate from benchmark leakage.
Screened from the abstract only, not tested. A reading guide, not a verification.
What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2
What the paper claims. Distilling GPT-2 into DistilGPT-2 destroys the indirect-object-identification circuit (35.2% vs 94.8% accuracy, no name-mover heads) while the greater-than circuit survives by moving its comparison head from layer 9 to layer 4.
Why it matters if true. If true, distillation keeps benchmark scores but can silently drop whole mechanisms, so a distilled model may fail where the teacher succeeded on similar tasks.
Explore deeper if. you ship distilled models or care what compression removes. Both models run on a CPU, so the main claim is cheap to recheck. Check first that the IOI accuracy metric is the same for both models.
What to doubt. Two circuits in two models only, and the IOI accuracy gap may reflect prompt format sensitivity in the small model rather than a missing mechanism.
Screened from the abstract, not tested. A reading guide, not a verification.
PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
What the paper claims. Adds a lightweight module that extracts physical-dynamics tokens from vision-language features so a robot policy understands dynamics without generating future video.
Why it matters if true. If true, robot policies get dynamics awareness without the latency of video world models.
Explore deeper if. you work on vision-language-action robot policies.
What to doubt. Robot results depend on benchmark and hardware; hard to verify without the code and robots.
Screened from the abstract only, not tested. A reading guide, not a verification.
Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs
What the paper claims. Diffusion vision-language models re-process the whole image at every denoising step although the image is rarely needed; a certificate-guided router skips it when not needed.
Why it matters if true. If true, inference cost of diffusion VLMs drops without discarding visual information.
Explore deeper if. you deploy diffusion language models with images.
What to doubt. The certificate is checked against one model family (LLaDA-V) in the abstract.
Screened from the abstract only, not tested. A reading guide, not a verification.
Hy-Nomas: Towards Free-form Interactive World Generation
What the paper claims. An 8B diffusion world model that follows successive text and camera commands while keeping history, using structured captions aligned with action-aware causal attention.
Why it matters if true. If true, interactive world generation gets better memory of earlier actions.
Explore deeper if. you build interactive world models with GPUs.
What to doubt. Video quality claims need demos; not toy-testable.
Screened from the abstract only, not tested. A reading guide, not a verification.
Mechanistic Insights into LoRA: Layer Sparsity for Adaptive Fine-Tuning via Path Patching
What the paper claims. LoRA's effect is concentrated in a few layers that differ by task, so selecting layers by zero-ablation lets Adaptive-LoRA beat all-layer LoRA with 30% of the layers.
Why it matters if true. If true, adapters can be cut by about 3.5x with no loss, and the layers chosen say where each skill lives.
Explore deeper if. you use LoRA and want smaller adapters, or study where tasks live in weights. Check first that the selected layers transfer to a different seed.
What to doubt. One model (Qwen2.5-1.5B), reasoning tasks only, and calibration data may overlap with evaluation.
Screened from the abstract, not tested. A reading guide, not a verification.
The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents
What the paper claims. A framework where multimodal agents recursively improve by sharing capability requirements across tasks of different modality, building environments and harness updates automatically.
Why it matters if true. If true, agent improvement could scale without hand-built environments and manual harness tuning.
Explore deeper if. you build self-improving agents and can access large compute.
What to doubt. Self-improvement claims are hard to separate from benchmark selection; large-scale and not toy-testable.
Screened from the abstract only, not tested. A reading guide, not a verification.
Work Agents: Systems, Capabilities, and Optimization
What the paper claims. A system-level perspective on work agents as foundation models plus runtime harnesses executing tasks in stateful workspaces.
Why it matters if true. A framing and survey for agent builders.
Explore deeper if. you build workflow agents.
What to doubt. Conceptual survey; no new tested claim in the abstract.
Screened from the abstract only, not tested. A reading guide, not a verification.
StreamVibe: Proactive Video Agents for Evolving User Intent
What the paper claims. Video agents that proactively respond to implicit, evolving user intent instead of waiting for explicit requests.
Why it matters if true. If true, assistants can help in real time without being asked.
Explore deeper if. you build streaming video assistants.
What to doubt. Intent benchmarks are subjective.
Screened from the abstract only, not tested. A reading guide, not a verification.
Interpretable sparse fine-tuning with TopKLoRA enables faithful circuit discovery
What the paper claims. TopKLoRA, a LoRA whose adapter activates only its top-k latents per token, lets researchers find and remove a backdoor circuit that is smaller than the dense LoRA equivalent without hurting instruction following.
Why it matters if true. If true, fine-tunes could be built to be inspectable and surgically fixable, which helps with poisoned training data.
Explore deeper if. you work on backdoors or weights you can audit. Check first against a plain LoRA with a sparsity penalty.
What to doubt. Backdoors are model organisms the authors created, which may be easier to localize than real ones.
Screened from the abstract, not tested. A reading guide, not a verification.
A Mathematical Theory of the Jacobian Lens
What the paper claims. A mathematical theory says the Jacobian lens finds directions that behave like leading Lyapunov modes (amplified across layers, stable across positions and contexts), validated on a Bayes filter and a toy transformer.
Why it matters if true. If true, it explains why the J-lens finds working-memory-like directions and gives a way to predict what other lenses will see.
Explore deeper if. you do interpretability theory and like toy systems with known ground truth. Check first that the toy transformer result is not specific to the hidden Markov setup used.
What to doubt. The paper is agnostic about the Global Workspace link it builds on, and validation is on toy systems, not frontier models.
Screened from the abstract, not tested. A reading guide, not a verification.
Nested Sparsity for Scalable Pretraining and Continual Learning
What the paper claims. Nestor, a transformer with a sparse-attention operator nested with mixture-of-experts, beats a conventional MoE at equal compute and sparsity and forgets less in continual pretraining.
Why it matters if true. If true, attention gets conditional computation too, not just the MLP.
Explore deeper if. you design architectures. Check first for equal-parameter, not just equal-compute, comparisons.
What to doubt. Pretraining claims need large compute to confirm.
Screened from the abstract, not tested. A reading guide, not a verification.
IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems
What the paper claims. A 320B-parameter mixture-of-experts coding-agent model with a documented pre-training, mid-training and post-training recipe.
Why it matters if true. If true, an open recipe for strong agentic CLI models.
Explore deeper if. you have large-scale training compute.
What to doubt. Model report; not testable at toy scale.
Screened from the abstract only, not tested. A reading guide, not a verification.
Cache Engineering for Agents
What the paper claims. Agent trajectories can be managed to keep key-value caches valid instead of trimming or summarising the prefix, which would invalidate later cache states.
Why it matters if true. If true, long-running agents cost less to serve.
Explore deeper if. you serve long-trajectory agents and care about KV cache cost.
What to doubt. Serving-cost savings depend on workload and infrastructure.
Screened from the abstract only, not tested. A reading guide, not a verification.
Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds
What the paper claims. A playable-world system that separates game execution (a light engine and code agent) from visual generation by a video model.
Why it matters if true. If true, game worlds keep persistent state and rules while generated by video models.
Explore deeper if. you build generated games.
What to doubt. System report; claims rely on demos.
Screened from the abstract only, not tested. A reading guide, not a verification.
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
What the paper claims. A controlled study of continued pre-training across data volume, diversity and model size asks whether scaling gains for embodied agents in physical settings carry over to varied virtual worlds.
Why it matters if true. If it holds, game data becomes a usable source of pretraining for embodied agents.
Explore deeper if. you train embodied or game agents.
What to doubt. Large-scale study; not testable at toy scale.
Screened from the abstract only, not tested. A reading guide, not a verification.
Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition
What the paper claims. A 35B mixture-of-experts agent that organises work, verifies and revises during long tasks.
Why it matters if true. If true, an open agent for long scientific tasks.
Explore deeper if. you have compute and build science agents.
What to doubt. Model report on the authors' benchmarks.
Screened from the abstract only, not tested. A reading guide, not a verification.
MemT-1.0: Towards Memory-Native Evolution
What the paper claims. A memory foundation model that builds memory into the architecture with a gated delta-rule layer and a mixture of storage, so it accumulates information across turns.
Why it matters if true. If true, long-term memory becomes part of model computation rather than an external store.
Explore deeper if. you study long-context or memory architectures.
What to doubt. Technical-report style, with claims by the authors on their own evaluations.
Screened from the abstract only, not tested. A reading guide, not a verification.
Learning from Fleet Experience: Collective Experience Learning Across 96 Robots
What the paper claims. Robots in a fleet pool their own execution experience to update a shared policy, instead of relying on expert corrections.
Why it matters if true. If true, deployed fleets can improve policies at scale without expert supervision.
Explore deeper if. you run robot fleets.
What to doubt. Real-robot results on 96 robots; cannot be reproduced without the fleet.
Screened from the abstract only, not tested. A reading guide, not a verification.
Deep Search in Real-World Web Environments
What the paper claims. A benchmark and task for deep-search agents on the real open web with mixed sources and interactions.
Why it matters if true. A more realistic test for research agents.
Explore deeper if. you build or evaluate web research agents.
What to doubt. Benchmark papers measure what they chose to include; web results drift over time.
Screened from the abstract only, not tested. A reading guide, not a verification.
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
What the paper claims. A harness giving multimodal models real-world visual tools and a multimodal memory, since raising the base model's inference budget alone gave modest gains on 22 benchmarks.
Why it matters if true. If true, tool use and memory beat extra compute for evidence-hungry visual tasks.
Explore deeper if. you build visual agents.
What to doubt. Results depend on the tool set and benchmarks chosen.
Screened from the abstract only, not tested. A reading guide, not a verification.
SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?
What the paper claims. A benchmark of 80 indoor scenes testing whether VLM agents can actively explore and build accurate metric 3D maps of objects.
Why it matters if true. If true, reveals whether spatial question answering reflects real spatial mapping.
Explore deeper if. you evaluate spatial agents.
What to doubt. A benchmark; results show gaps, not a method.
Screened from the abstract only, not tested. A reading guide, not a verification.
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
What the paper claims. A pipeline that synthesises controllable environments and tasks for training computer-use agents, avoiding security risks of real systems.
Why it matters if true. If true, scalable and safe training for computer-use agents.
Explore deeper if. you train computer-use agents.
What to doubt. Synthetic environments may not match real ones.
Screened from the abstract only, not tested. A reading guide, not a verification.
Learning When to Abort in Transformer Planning: A Mechanistic Study
What the paper claims. Transformers trained to emit an abort token when they generate an invalid graph transition plan better; the paper proves a width lower bound for the detector and studies the mechanism.
Why it matters if true. If true, teaching error detection helps planning capacity, and the width bound says how big a network must be to catch all invalid steps.
Explore deeper if. you study planning in small transformers; the path-finding setting is cheap to rebuild on CPU.
What to doubt. Synthetic path-finding on fixed graphs; unclear how far it transfers to language-model planning.
Screened from the abstract only, not tested. A reading guide, not a verification.
The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?
What the paper claims. Whether decoding individual ViT patch tokens gives readable output is not an inherent transformer property; it varies with architecture, scale and training objective.
Why it matters if true. If true, interpretability diagnostics built on token decoding cannot be assumed to work on a new model.
Explore deeper if. you use patch-token decoding as a diagnostic in vision models.
What to doubt. Abstract is a systematic survey; the causes of the variation are described as implicit biases, not isolated.
Screened from the abstract only, not tested. A reading guide, not a verification.
Anisotropic Representation Dynamics in Transformers
What the paper claims. A theory of when an attention layer turns a small difference between two inputs, which would otherwise shrink, into one that grows. It gives an exact spectral test for whether this switching happens and which directions it affects, checked on four transformer architectures.
Why it matters if true. Relevant to robustness and sensitivity: it says when changes in attention routing amplify perturbations, and in which directions.
Explore deeper if. you study adversarial sensitivity, representation stability or the mathematical theory of attention.
What to doubt. A local, single-layer analysis; the abstract does not show that the identified directions matter for any downstream behaviour.
Screened from the abstract, not tested. A reading guide, not a verification.
BMuon: Budgeting Muon's Spectral Geometry for Structured Compressibility
What the paper claims. Transformers trained with the Muon optimizer fall apart when you remove MLP channels after training (a vision model drops from 75% to 3% accuracy at 30% removal). The paper proposes BMuon, a modified Muon that makes hidden channels compete for a shared budget so that pruning afterwards does little damage.
Why it matters if true. Muon is becoming a default optimizer for large-model pretraining, and structured pruning is a standard way to shrink models for deployment. If Muon models resist pruning, teams have to choose between the two or use a variant like BMuon.
Explore deeper if. you train with Muon and plan to prune or compress afterwards, or you study how optimizer geometry shapes what a network learns. Check first whether the collapse holds against an AdamW baseline and with a pruning rule other than weight magnitude.
What to doubt. The pruning comparisons include no AdamW baseline, so it is not shown that Muon is worse than the usual optimizer. Channels are ranked by weight magnitude, and Muon's updates tend to equalize magnitudes, so the collapse may say more about the ranking rule than about the network. BMuon also gives up dense quality (perplexity 12.84 vs 10.92 at 1B).
Screened from the HTML full text, not tested. A reading guide, not a verification.
HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
What the paper claims. A coding agent controls a two-arm robot through a command line, writing and fixing small vision tools (camera triangulation) after each failure. It sets a new state of the art on 42 robot tasks (39.7% success). It is excellent at pushing (92.5%) but poor at tight insertion (8.3%): the limit is contact, not reasoning.
Why it matters if true. Shows how far plain coding agents get in robotics and pinpoints where they stop, which tells the field where learned policies are still needed.
Explore deeper if. you combine coding agents with robots, or want a clear split between reasoning and contact-control failures. Tools are open-sourced.
What to doubt. Tools were built on ten layouts of one seed per task and tested on layouts of the same seed, which is a limited test of generalization.
Screened from the abstract, not tested. A reading guide, not a verification.
MOSS-Image Technical Report
What the paper claims. An open 5B diffusion image model aimed at dense text and complex layouts, with intent-aware composition.
Why it matters if true. If true, an open-weight option narrows the gap on text-heavy images.
Explore deeper if. you generate posters or infographics with open models.
What to doubt. Technical report on the authors' own evaluations.
Screened from the abstract only, not tested. A reading guide, not a verification.
Spurious Attention Heads in Vision Transformers: Mechanistic Identification of Annotation-Driven Bias in Medical Chest X-Ray Analysis
What the paper claims. In ViT models for chest X-rays, a single attention head reads the text annotations in image corners instead of looking at the lungs (62-88% of images across datasets and models). Fine-tuning does not remove it, and ablating the heads leaves output metrics unchanged, so standard evaluation cannot detect it.
Why it matters if true. A concrete case where a shortcut is real and causal yet invisible to output testing, a warning for medical AI validation.
Explore deeper if. you validate medical imaging models or audit vision models for shortcuts.
What to doubt. Main analysis on 200 COVID images; that ablation changes nothing at the output makes the head's clinical impact hard to quantify.
Screened from the abstract, not tested. A reading guide, not a verification.
Towards Continual Learning: Mitigating Forgetting With Subspace Tuning
What the paper claims. Fine-tuning causes forgetting mostly by changing a weight matrix along its strongest directions. Subspace Tuning (SST) freezes the top 99% of each matrix's singular directions and lets the optimizer change only the weakest 1%. It keeps old skills almost intact on 8B-14B models while learning the new task about as well.
Why it matters if true. Forgetting is the main cost of fine-tuning a model on a new domain. If a simple projection of each optimizer step nearly removes it, with no adapters, replay data or extra memory, it could become a default fine-tuning setting.
Explore deeper if. you fine-tune pretrained models and lose general ability, or you study where forgetting lives in weight space. The method is a few lines around any optimizer, so it is cheap to try.
What to doubt. Learning rates were not tuned separately for SST and plain fine-tuning, so the comparison is not at matched learning; the authors list full learning-retention frontiers as future work. The benefit shrinks at small scale (the 1.7B model still loses 8 points), it did not help on one instruct model, and no code release is mentioned.
Screened from the HTML full text, not tested. A reading guide, not a verification.
Translation Circuit Discovery and Targeted Optimization for Low-Resource Multilingual Translation
What the paper claims. Finds the translation circuits for Hindi-Urdu and Hindi-Sindhi in multilingual LLMs and fine-tunes only those (under 5% of parameters) with preference optimization. BLEU improves 34-46% over the strongest baselines, and translation appears to run through a shared semantic space rather than via English.
Why it matters if true. An example of interpretability guiding efficient adaptation for low-resource languages.
Explore deeper if. you work on low-resource translation or circuit-guided fine-tuning.
What to doubt. Two language pairs; the large relative BLEU gains depend on weak baselines for these languages.
Screened from the abstract, not tested. A reading guide, not a verification.
VALCON: Benchmarking and Enhancing Value Consistency in Large Language Models via Dual-Role Variation
What the paper claims. Across 28 LLMs, models often judge their own outputs inconsistently with the values they follow when generating. Probing shows the generator and evaluator roles use decoupled representations, and a representation-level intervention reduces the inconsistency.
Why it matters if true. Many alignment pipelines use the same model to act and to judge; if the roles disagree, self-evaluation is less trustworthy than assumed.
Explore deeper if. you use LLMs as judges of their own or similar outputs, or build self-critique pipelines. Code and data are released.
What to doubt. Benchmarks are synthesized by the authors' framework, so the measured inconsistency depends on its design.
Screened from the abstract, not tested. A reading guide, not a verification.
Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness
What the paper claims. Measures uncertainty as how small an internal intervention flips the model's answer. Wrong and uncertain answers flip with smaller pushes, the signal complements output-based uncertainty, and facts learned correctly become harder to flip after training.
Why it matters if true. Offers an uncertainty signal grounded in internal mechanisms, which may catch confident-sounding hallucinations that output probabilities miss.
Explore deeper if. you need hallucination detection or uncertainty estimates for open models.
What to doubt. Three models and QA datasets; computing the margin requires internal access and extra computation.
Screened from the abstract, not tested. A reading guide, not a verification.
Sparse Autoencoders Can Induce Local Geometry for Mitigating Catastrophic Forgetting
What the paper claims. Trains a sparse autoencoder at each task boundary and turns its local geometry into a penalty on how representations may drift when learning the next task. With replay, it beats standard continual-learning baselines on pretrained ViT benchmarks by 3-12 points.
Why it matters if true. Gives a principled way to decide which representation changes to forbid while still allowing learning, and the constraints can be kept without keeping the SAE.
Explore deeper if. you do class-incremental continual learning with a pretrained backbone and already use replay.
What to doubt. Gains are on pretrained ViT-B/16 with replay. The authors report mixed results when training from scratch, and comparisons are against classic baselines rather than recent PEFT-based methods.
Screened from the abstract, not tested. A reading guide, not a verification.
Representation Redistribution in Sparse Autoencoders
What the paper claims. Rewording a prompt without changing its meaning shifts activity between sparse-autoencoder features. The paper scores which feature changes recur consistently across rewordings, and separates such representation changes from whether they actually change a downstream decision; about 20 features predict safety-readout failures almost as well as the full set.
Why it matters if true. Warns against reading every SAE feature change as meaningful, and gives a way to find the changes that recur and matter, which is useful for safety monitors built on SAE features.
Explore deeper if. you build safety or behaviour monitors on SAE features, or study how stable SAE features are under paraphrase. Code is released.
What to doubt. Results are on prompt transformations and two linear safety readouts chosen by the authors; whether the score transfers to other readouts or nonlinear decisions is not shown in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation
What the paper claims. When distilling what RL taught a teacher model into a student, compare the teacher's change early in RL with its change late in RL. Transfer the final change only where the two agree in direction. This beats endpoint-only distillation in all six math and code settings tested.
Why it matters if true. RL is expensive to run on every model size. A cheap way to copy RL's effect from a small trained teacher into bigger students, costing one extra scoring pass, could save a lot of compute.
Explore deeper if. you run RL on one model and want to transfer the gains to other model sizes, or you study what RL changes in a policy.
What to doubt. One teacher (Qwen3-4B), one model family and two domains; it needs an intermediate RL checkpoint, which not every pipeline keeps.
Screened from the abstract, not tested. A reading guide, not a verification.
FDE-Gym: The Last Mile of AI Delivery
What the paper claims. A benchmark of 95 realistic customer-delivery tasks where agents must clarify vague requirements with simulated stakeholders and work in live-like systems. The best agent fully resolves only 14 of 95. Discovering unstated constraints is the main bottleneck, but even with all facts the agent often fails.
Why it matters if true. Measures the gap between solving well-specified coding tasks and delivering working systems for real customers, which is the job companies actually want agents to do.
Explore deeper if. you build or deploy agents for client-facing engineering work, or need an evaluation beyond issue-resolution benchmarks.
What to doubt. Stakeholders are scripted; 95 tasks is small, and scores depend on how the simulation encodes hidden constraints.
Screened from the abstract, not tested. A reading guide, not a verification.
Has GPT-6-Astra Achieved Embodied Intelligence?
What the paper claims. A broad evaluation of GPT-6-Astra on 12 robot benchmarks. With in-context demonstrations and longer execution it performs at an unprecedented level, but on 8 real robots the demonstrations are expensive and taking longer is not physically feasible. The conclusion: competent, not yet practical.
Why it matters if true. A reality check on whether general frontier models can replace specialized robot policies, with concrete gaps listed.
Explore deeper if. you decide between general models and specialized policies for robot control, or design embodied benchmarks.
What to doubt. Simulation results use a protocol with extra steps and in-domain demonstrations that real deployments may not allow, which the authors point out.
Screened from the abstract, not tested. A reading guide, not a verification.
Behavioral Compositionality of Sparse Autoencoder Representations
What the paper claims. Measures how far a model's behavior can be explained by adding up the effects of individual sparse-autoencoder features. Learned SAE coordinates are more additive than random rotations of the same subspace in all 18 settings, but not fully additive: pairwise interaction terms clearly improve predictions, including for combined steering.
Why it matters if true. Steering and explanation with SAEs usually assume features act independently. This quantifies when that fails and shows pairwise terms fix much of it.
Explore deeper if. you steer with several SAE features at once or explain behavior as a sum of feature effects. Code is released.
What to doubt. Results depend on the sampled feature subsets and chosen behaviors; higher-order interactions beyond pairs are only partly examined.
Screened from the abstract, not tested. A reading guide, not a verification.
Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey of Recoverability, Interpretability, and Utility
What the paper claims. A critical survey of 138 sparse-autoencoder studies in NLP that organizes what is known about whether SAE features are recoverable, interpretable and useful, reconciles conflicting results, and lists open questions, with a public research register.
Why it matters if true. A map of a crowded field that separates well-supported SAE claims from weak ones; useful before building on any particular SAE result.
Explore deeper if. you are starting SAE work or need to judge a specific SAE claim against the rest of the literature.
What to doubt. A survey's conclusions depend on which studies were included and how they were compared; it reports no new experiments.
Screened from the abstract, not tested. A reading guide, not a verification.
Beyond State Capacity: Why Retrieval Breaks in Mamba
What the paper claims. Mamba fails at in-context retrieval even when the question comes before the context, so its fixed-size memory is not the limit. The selection gates could solve the task, but the signal saying which tokens matter appears only in deeper layers, after the layers where retrieval happens.
Why it matters if true. Changes the diagnosis of why state-space models trail transformers at retrieval: from 'not enough memory' to 'relevance arrives too late in depth'. That points to architectural fixes other than bigger states.
Explore deeper if. you design or deploy state-space or hybrid models and care about retrieval, or you do probing and causal interventions on recurrent models.
What to doubt. The abstract gives no model sizes or numbers, and it is unclear how far the result goes beyond the specific retrieval tasks and Mamba versions studied.
Screened from the abstract, not tested. A reading guide, not a verification.
Sparse Autoencoders Are a Shared Coordinate System for Symbolic Matching
What the paper claims. Uses sparse-autoencoder codes, reweighted by how rare each feature is, to match items across different symbolic systems (ontologies, taxonomies and similar). It beats the same model's dense embeddings on 12 tasks by up to 45 points, and each match can be explained by named features.
Why it matters if true. A practical, interpretable use of SAEs: matching entries across knowledge systems, a first step toward merging them automatically.
Explore deeper if. you align ontologies or knowledge graphs, or want a downstream use case where SAE features beat dense embeddings.
What to doubt. The baseline is the same LLM's dense embeddings; the abstract does not compare against strong dedicated matching systems.
Screened from the abstract, not tested. A reading guide, not a verification.
Complete or Sparse: A Tale of Two Identifiabilities
What the paper claims. Proves when a network trained only to classify will end up with neurons that each represent one true hidden factor of the data. If the hidden layer has exactly as many units as there are factors, activations are non-negative, and there is pressure toward small activations, each neuron encodes exactly one factor. With too few units, clean recovery generically fails.
Why it matters if true. Interpretability work assumes that features in a network line up with real factors. This gives exact conditions for when that is guaranteed, and a reason why 'grandmother-cell' neurons might appear in brains and in ReLU networks.
Explore deeper if. you work on sparse autoencoders, disentanglement or identifiability, or you want a clean synthetic test bed for whether a training setup recovers true latent variables. Code is released.
What to doubt. Needs exactly matched capacity, non-negative latents, equal-entropy classes and uniform class priors, which real data rarely meets. The authors report that alignment converges slowly and that the evidence on language models and monkey cortex is weak or not statistically significant.
Screened from the HTML full text, not tested. A reading guide, not a verification.
Training Dynamics of Transformers on Unbounded Inputs: Attention Learning Can Lag Output Layer Learning
What the paper claims. For a small transformer learning to recognize a simple formal language over inputs of any length, training must update the output layer first, then attention, then the output layer again. Training only attention, only the output layer, or attention first fails. The authors find the same 'attention lags the output layer' pattern helps on other tasks.
Why it matters if true. Challenges the theory picture where one gradient step learns the features and the readout is fitted afterwards. If attention needs a trained readout before it can learn, that changes how training dynamics of attention models should be analyzed and possibly scheduled.
Explore deeper if. you study the theory of transformer training dynamics, or want a small, provable setting to test learning-rate schedules per layer.
What to doubt. The proof is for a shallow model, one language and a stylized three-phase procedure, not ordinary joint training. The abstract gives no numbers for the 'other tasks' evidence.
Screened from the abstract, not tested. A reading guide, not a verification.
More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning
What the paper claims. A fair comparison of ten self-distillation methods for agent RL against plain GRPO, in 132 runs and over 12,000 GPU-hours across four environments and three model sizes. No method consistently beats GRPO; strong final scores can hide unstable training, and extra compute does not reliably help.
Why it matters if true. A reality check on a crowded method space; anyone choosing an agent-RL method should start from GRPO and these results.
Explore deeper if. you pick an RL recipe for training agents, or review agent-RL papers.
What to doubt. Four environments and ten methods; settings chosen for fairness may not match each method's best configuration.
Screened from the abstract, not tested. A reading guide, not a verification.
LEV: Thinking Fast and Deep with Looped Decision Models
What the paper claims. Decision models that answer in one forward pass can be made to think longer on hard questions by looping their layers, instead of generating more text.
Why it matters if true. If true, you get adaptive compute at inference for classifiers and decision heads without paying for decoded tokens.
Explore deeper if. you build fast classifiers or judges and want depth that adapts to difficulty, which can be tested on small looped transformers.
What to doubt. Gains from looping often depend on the task suite; the abstract does not show whether the extra depth is used only when needed.
Screened from the abstract only, not tested. A reading guide, not a verification.
BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
What the paper claims. A business-operations benchmark that scores the validity of intermediate agent steps under conflicting policies, not just final success.
Why it matters if true. If true, exposes agents that succeed with invalid steps in payments-style tasks.
Explore deeper if. you deploy agents in high-stakes workflows.
What to doubt. A benchmark with the authors' task design.
Screened from the abstract only, not tested. A reading guide, not a verification.
Neuro-Symbolic Graph Transformers for Hierarchical Intention Recognition from Multimodal Human Activity Sequences
What the paper claims. A graph transformer with a human-reasoning graph and a logic-based loss predicts actions, activities and two levels of intention from human motion data, improving logical consistency between levels at a small accuracy cost.
Why it matters if true. Robots collaborating with people need to infer goals, not just actions; this adds structure to that inference.
Explore deeper if. you work on human intention recognition for human-robot collaboration.
What to doubt. The hierarchy dataset is synthesized from an existing action dataset, so intention labels are derived, not observed.
Screened from the abstract, not tested. A reading guide, not a verification.
Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation
What the paper claims. Distilling long chain-of-thought works nearly as well when trained only on flawed reasoning traces, a narrow set of prompts, or problems the model already solves. What transfers is the reasoning behavior, not facts. Standard cross-entropy loss is needed to learn complex reasoning patterns.
Why it matters if true. Data curation for reasoning distillation can focus on the style of reasoning rather than correctness and coverage, which is much cheaper.
Explore deeper if. you build reasoning-distillation datasets. Code is released.
What to doubt. 'Nearly comparable' leaves some gap; results are for the models and tasks tested.
Screened from the abstract, not tested. A reading guide, not a verification.
DTDD: Divergence-Triggered Dynamic Distillation for Reliable On-Policy Supervision
What the paper claims. In on-policy distillation, when the student drifts too far from the teacher, let the teacher take over and generate a recovery, which the student then imitates. This stabilizes training (8.6x lower peak gradient norms) and reaches the same accuracy 2.6x faster, with gains of about 4-5 points in agent settings.
Why it matters if true. Makes on-policy distillation more stable when student and teacher start far apart, common in long reasoning and tool use.
Explore deeper if. you distill into weaker students or train agents with on-policy distillation.
What to doubt. Final accuracy is about the same as plain OPD in the math setting; the gains are in speed and stability.
Screened from the abstract, not tested. A reading guide, not a verification.
Confine to the Task, Anchor the Rest: Concept-Guided Retention Regularizer for Parameter-Efficient Fine-Tuning
What the paper claims. A regularizer for parameter-efficient fine-tuning that uses a frozen sparse autoencoder to name concepts. It keeps the update inside the concepts the new task needs, and holds a chosen set of general concepts at their original activations, which reduces forgetting at similar task accuracy.
Why it matters if true. Most anti-forgetting methods constrain only where the weights move; this also constrains what the activations do. Because the protected directions are named concepts, people can inspect and choose what is protected.
Explore deeper if. you fine-tune with LoRA or similar and have, or can train, an SAE for the base model, and you want control over which abilities are preserved.
What to doubt. Depends on an SAE being available and its concepts being faithful. Results are on CLIP and Llama-3.1-8B only; the abstract does not say how the general-purpose concept set is selected or how sensitive results are to that choice.
Screened from the abstract, not tested. A reading guide, not a verification.
Does self-distillation improve reasoning?
What the paper claims. Re-examines two self-distillation methods for reasoning that reported beating reinforcement learning. Their gains mostly do not carry over to unseen problems, are explained by overfitting or by reusing reasoning the model already had, and their training directions agree only weakly with directly optimizing reward.
Why it matters if true. A replication-style check on a fast-moving claim: anyone choosing self-distillation over RL for reasoning should know the reported gains may be in-distribution only.
Explore deeper if. you train reasoning models with self-distillation or privileged information, or you compare post-training methods. Code is released.
What to doubt. Two methods and models from 1.7B to 14B; it does not rule out that other self-distillation designs work.
Screened from the abstract, not tested. A reading guide, not a verification.
Agent-Reflex: Teaching Language Agents to Act Reflectively
What the paper claims. A training pipeline that teaches tool-using agents to reflect and recover when tools return noisy results or users correct them: mine success/failure pairs into reflection skills, fine-tune on them, then combine RL with reflection-guided distillation. Qwen3-8B improves on long-horizon tool-use and noisy-interaction benchmarks.
Why it matters if true. Real tool services are noisy and users interrupt; agents trained only on clean environments break. This targets robustness directly.
Explore deeper if. you train tool-using agents that face unreliable APIs or user corrections. Code is released.
What to doubt. Three-stage pipeline with many parts; the abstract does not say which stage contributes most.
Screened from the abstract, not tested. A reading guide, not a verification.
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features
What the paper claims. Tests whether attribution methods such as SHAP and Integrated Gradients correctly say which sparse-autoencoder features drive a GPT-2 sentiment classifier, by comparing them with actually removing or steering each feature. Most agreements vanish after multiple-comparison correction, and methods that agree with each other still disagree with the measured causal effects.
Why it matters if true. Many interpretability claims rest on attribution scores. If methods agree with each other but not with interventions, agreement between them is not evidence of correctness, and causal checks are needed.
Explore deeper if. you rank SAE features by attribution, or use agreement between explanation methods as validation.
What to doubt. One model (GPT-2 Small), one layer, one SAE and one task; whether the conclusions carry to other models and tasks is untested.
Screened from the abstract, not tested. A reading guide, not a verification.
OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding
What the paper claims. A benchmark of 820 videos testing when something happens, where it is, and what happens, with an audio-visual subset. Across 50 multimodal models, open models trail proprietary ones (especially with audio), precise time and location grounding stays hard, and answers change depending on whether the target is described in text or with an image.
Why it matters if true. Gives a single diagnostic for video understanding that separates grounding from question answering, useful for choosing or improving video models.
Explore deeper if. you build or pick video-language models, especially for tasks needing precise temporal or spatial grounding or audio.
What to doubt. A benchmark of 2,460 instances; conclusions depend on its coverage and annotation choices, and rankings can shift as models are updated.
Screened from the abstract, not tested. A reading guide, not a verification.
Self-Evolving Agent Harness
What the paper claims. A 66-page framework and theory for agents that keep revising their own harness (prompts, memory, tools, code) across runs. It gives generalization and retention bounds under stated assumptions and argues that safe deployment needs fresh validation data and independent safeguards.
Why it matters if true. Self-modifying harnesses are being deployed; this offers vocabulary and guarantees for when accepted changes can be trusted not to regress.
Explore deeper if. you design or govern systems where agents modify their own tooling, or want a reference treatment of harness self-improvement.
What to doubt. Mostly conceptual and theoretical; guarantees rest on assumptions (independent validation data, regression budgets) that may be hard to meet in practice.
Screened from the abstract, not tested. A reading guide, not a verification.
Into the Unknown: Benchmarking AI agents on open scientific research problems
What the paper claims. Scientists at US national labs each gave an unpublished follow-up direction to their own work, with conditions a real follow-up must meet. Nine frontier agents wrote 135 follow-up papers: none met all conditions, and 114 failed an integrity check, mostly by claiming work never done or reporting numbers that contradict their own logs.
Why it matters if true. A sober measurement of AI-scientist claims on genuinely open problems, and a warning that fabricated results are the common failure mode.
Explore deeper if. you build or evaluate research agents, or consider using agents to draft scientific work. Code and data are released.
What to doubt. Fifteen problems from one institution type and five domains; the integrity check and rubrics are authored by the problem contributors.
Screened from the abstract, not tested. A reading guide, not a verification.
Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
What the paper claims. A dataset of 1,170 long-horizon agent environments built from a hidden full specification, which gives a natural measure of progress at every step. Turning progress into dense rewards makes hard tasks trainable; the resulting 27B model matches frontier models at pass@5 on one benchmark and generalizes to terminal tasks.
Why it matters if true. Sparse rewards and missing data are the main obstacles to training open long-horizon agents; this addresses both.
Explore deeper if. you train open models for multi-hour agent tasks and need dense reward signals.
What to doubt. Comparisons at pass@5 flatter the trained model; generalization is shown on a few benchmarks.
Screened from the abstract, not tested. A reading guide, not a verification.
GreenWAM: A World Foundation Model for Robotics
What the paper claims. One robot world model trained on 20 robot families through a shared joint-coordinate action space works both as a world simulator and as a policy. It is top-tier on world-model benchmarks and competitive as a policy, and beats adapted baselines on a real humanoid.
Why it matters if true. A single model serving both roles across many robot bodies would simplify robot learning pipelines and data sharing.
Explore deeper if. you train across robot embodiments or want one model for simulation and control.
What to doubt. Policy results are 'competitive' rather than best on several benchmarks; the shared action space may not fit unusual robots.
Screened from the abstract, not tested. A reading guide, not a verification.
Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents
What the paper claims. Instead of summarizing a long agent's history, keep context organized around reusable task objects that are updated when evidence changes, plus recent interactions. Improves success by 6-7% on computer-use and browsing benchmarks while cutting tokens by 23-52%.
Why it matters if true. Long agents spend most tokens re-reading their history; organizing by progress can make them cheaper and better.
Explore deeper if. you run long-horizon computer-use or browsing agents and pay for large contexts.
What to doubt. Tested on two benchmarks; designing the right task objects may be domain-specific.
Screened from the abstract, not tested. A reading guide, not a verification.
Making LoRA Identifiable: Orthogonal Alignment for Continual Learning
What the paper claims. Keeps a single LoRA adapter across a sequence of tasks without growing memory. It constrains one LoRA matrix to orthonormal columns and then rotates the other matrix to line up with what earlier tasks learned, merging only the directions where old and new tasks agree.
Why it matters if true. Most continual-LoRA methods add a new adapter per task, so memory grows with the number of tasks. A fixed-size adapter that still avoids forgetting would make long task sequences practical.
Explore deeper if. you run continual fine-tuning with LoRA and need memory to stay flat as tasks accumulate.
What to doubt. The abstract gives no numbers, model sizes or forgetting margins, and does not say how much the orthonormal constraint costs on any single task.
Screened from the abstract, not tested. A reading guide, not a verification.
In-context classification: Calibration and noise in linear attention with softmax readout
What the paper claims. An exact high-dimensional theory of in-context binary classification in a linear-attention model with a softmax output. It separates how calibration, noise in the context examples and estimation noise each affect the prediction, and shows how context length, pretraining size, regularization and temperature trade off.
Why it matters if true. Gives closed-form predictions for when in-context learning is well calibrated, which could guide prompt length and pretraining choices in simple settings.
Explore deeper if. you work on the theory of in-context learning or on calibration, and want formulas you can check against small simulations.
What to doubt. Linear attention in a proportional high-dimensional limit; the softmax-attention results are described only as qualitatively related.
Screened from the abstract, not tested. A reading guide, not a verification.
One-Layer Transformers Provably Learn In-Context K-Nearest Neighbor Prediction with Chain-of-Thought
What the paper claims. A one-layer transformer provably cannot do k-nearest-neighbour prediction for k > 1 in a single step, but can with chain-of-thought, and gradient descent can learn the chain-of-thought solution in a stylized setup.
Why it matters if true. A concrete, provable example of chain-of-thought adding capability that a single forward pass lacks, rather than only adding compute.
Explore deeper if. you study why chain-of-thought helps or the expressivity of small transformers, and want a clean task where the gap is proven.
What to doubt. The negative result is for odd k on a well-separated binary task, and the learning result uses a stylized training setup; how far it extends to ordinary training is open.
Screened from the abstract, not tested. A reading guide, not a verification.
TPPO: Tracking-Value PPO For Language Model Reasoning
What the paper claims. A new value model for PPO in language-model reasoning, built on the idea that value updates add up in log-odds along the response. It adds 7-8 points on math over standard PPO and generalizes better, with gains on other tasks.
Why it matters if true. Better token-level credit assignment could make PPO competitive again with critic-free methods like GRPO.
Explore deeper if. you train reasoning models with PPO or study critics for sparse rewards. Code is released.
What to doubt. Compared with PPO; the abstract does not compare with GRPO-style methods most teams now use.
Screened from the abstract, not tested. A reading guide, not a verification.
Vela: User Behavior as Supervision Sparks LLM Intelligence
What the paper claims. Turns real behavior records from 100,000 users into about 75,000 training tasks. Fine-tuning on them improves user-related tasks by up to 11.5 points and general benchmarks by about 4 points, and more behavior data keeps helping.
Why it matters if true. Suggests user behavior logs are a scalable training signal not only for personalization but for general ability, which matters to platforms that hold such data.
Explore deeper if. you build personalization, recommendation or assistant models with access to behavior logs.
What to doubt. Data is from one platform's users; privacy and representativeness of such data are serious questions not addressed in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Continually Post-training Small Models via Synthetic Self-distillation
What the paper claims. Small models forget old skills more, and learn new ones worse, the further fine-tuning pulls their outputs away from how they naturally write. The fix: have a slightly larger model from the same family reword the training answers into the small model's own style, keep only rewordings that are still correct, and train on those.
Why it matters if true. Anyone updating a small or on-device model with new knowledge faces forgetting. If this holds, a cheap data-side change (rewording targets) protects old skills better than six existing methods, without changing the training algorithm.
Explore deeper if. you fine-tune small models repeatedly, consolidate new facts into adapters, or see unrelated knowledge degrade after a narrow fine-tune. It is also a natural explanation to test whenever a fine-tune leaks into neighbouring facts.
What to doubt. The abstract says forgetting correlates 'linearly' with distribution shift, but the paper reports only a rank correlation, which shows ordering, not a straight line. It needs a larger same-family model to do the rewriting, and no code release is mentioned.
Screened from the HTML full text, not tested. A reading guide, not a verification.
Time Delay System for Sequence Modeling
What the paper claims. A new sequence model that recalls past inputs by how similar they are to the current one, using a time-delay formulation with constant cost per step and interpretable rules.
Why it matters if true. If it holds, it is an efficient alternative to attention with explicit, inspectable memory recall.
Explore deeper if. you work on efficient sequence architectures or interpretable memory mechanisms.
What to doubt. The abstract claims state-of-the-art results on classification and generation without naming benchmarks or scale.
Screened from the abstract, not tested. A reading guide, not a verification.
Language Models Have Interpretable Narrative Mechanisms
What the paper claims. Language models contain specific neurons, attention heads and learned features linked to how story-like a text is. Switching them off hurts prediction on story-like text more than on other text, and makes generations less narrative.
Why it matters if true. It is causal evidence that a fairly abstract text property has identifiable machinery inside models, which matters for interpretability and for steering story generation.
Explore deeper if. you study interpretability of high-level text properties, or build story-generation or narrative-QA systems.
What to doubt. Effects vary across models and languages, results on hallucination benchmarks are mixed, and the authors say the organisation of these components remains an open question.
Screened from the abstract, not tested. A reading guide, not a verification.
Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers
What the paper claims. An attention design that runs linear-cost attention while reading the prompt and full attention while generating, claiming a 21.9x prefill speedup with near-perfect long-context retrieval.
Why it matters if true. Long-prompt, short-answer workloads are common; a big prefill speedup without losing recall would cut serving cost.
Explore deeper if. you serve long-context models or convert existing transformers to efficient attention.
What to doubt. Strong 'successor to the Transformer' framing; needle-in-a-haystack is a weak test of real long-context ability.
Screened from the abstract, not tested. A reading guide, not a verification.
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels
What the paper claims. When softmax attention is replaced with a cheaper linear approximation, most of the error comes from a few unusually large dimensions in the queries and keys. Handling those few dimensions separately makes the approximation noticeably more accurate.
Why it matters if true. Converting a pretrained model to linear-time attention is a route to cheaper long-context inference. If error concentrates in a few dimensions, conversion methods should treat them specially.
Explore deeper if. you convert pretrained transformers to linear or efficient attention, or study outlier dimensions. It was measured on Qwen2.5-0.5B, so the outlier analysis is easy to repeat, and code is released.
What to doubt. Gains are modest (about 9% lower error than one baseline) and shown on a single small model; the perplexity comparison is against one learned baseline.
Screened from the abstract, not tested. A reading guide, not a verification.
In-Context Learning for System Identification: Probabilistic Formulation, Long-Sequence Handling, and Benchmark Validation
What the paper claims. A pretrained transformer can identify dynamical systems from example data in-context, with calibrated uncertainty and long-sequence handling, and can generate synthetic data to help classical methods.
Why it matters if true. Control and system-identification engineers could skip per-system model fitting.
Explore deeper if. you do system identification or control and have little data per system.
What to doubt. Evaluated on standard benchmarks with 'competitive' accuracy; benefits concentrate in data-scarce and shifted settings.
Screened from the abstract, not tested. A reading guide, not a verification.
PDE-PFN: Prior-Data Fitted Neural PDE Solver
What the paper claims. A transformer pretrained on noisy approximate PDE solutions can predict new PDE solutions in-context, without gradient updates, across different equations.
Why it matters if true. Fast surrogate solvers for scientific computing that adapt from examples instead of retraining.
Explore deeper if. you build neural PDE solvers or physics surrogates.
What to doubt. Two-dimensional PDEs only; accuracy is 'competitive or improved' against the evaluated baselines.
Screened from the abstract, not tested. A reading guide, not a verification.
Exploring Attention Sinks in Text-to-Image Transformers from Diffusion and Autoregressive Models
What the paper claims. Attention sinks (positions that soak up attention regardless of meaning) also appear in text-to-image transformers, but in different places for diffusion models (modality boundaries, image corners) than for autoregressive ones (the first tokens).
Why it matters if true. Sinks affect efficiency tricks and quantization in language models; knowing where they sit in image generators guides similar engineering there.
Explore deeper if. you optimize, compress or interpret text-to-image transformers.
What to doubt. Perturbing sink positions sometimes had little effect, so their functional importance is mixed; findings are architecture-specific.
Screened from the abstract, not tested. A reading guide, not a verification.
Scaling Agentic Data for Long-Horizon Terminal Intelligence
What the paper claims. An automated engine that builds long, realistic terminal tasks from real professional work, checking that the environment has not already done the hard parts. Fine-tuning a 27B model on 1,500 such environments improves all 12 coding and agent benchmarks.
Why it matters if true. Long-horizon agent training data is scarce; this offers a way to generate it while avoiding environments that make tasks too easy.
Explore deeper if. you train terminal or coding agents and need long-horizon trajectories.
What to doubt. Gains are from supervised fine-tuning on the authors' data; the 'self-improvement' case studies are illustrative.
Screened from the abstract, not tested. A reading guide, not a verification.
On the Design Space of Continuous Diffusion Language Models
What the paper claims. Explains why 'self-conditioning' helps continuous diffusion language models (it fixes a mismatch between training and generation) and proposes cheaper variants that match discrete diffusion on a math benchmark.
Why it matters if true. Useful design guidance for diffusion-based language models, which promise faster and more controllable generation.
Explore deeper if. you build or study diffusion language models.
What to doubt. Main gains are on one small math benchmark; the 1.7B comparison wins four of six likelihood evaluations.
Screened from the abstract, not tested. A reading guide, not a verification.
Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers
What the paper claims. Generalizes multi-head attention so heads can share similarity information through a learned, input-dependent metric, claiming better accuracy at similar cost.
Why it matters if true. A drop-in attention upgrade, if real, would be easy to adopt.
Explore deeper if. you experiment with attention variants on your own benchmarks; code is released.
What to doubt. Benchmark-win framing with no named scale in the abstract; extra parameters may explain part of the gain.
Screened from the abstract, not tested. A reading guide, not a verification.
Not Just an Evaluation: A Case Study of Discourse Particle Sensitivity in Language Models
What the paper claims. Language models, like people, are sensitive to small discourse words like 'just' and 'only', and instruction tuning increases that sensitivity. Inside the models, units for 'only' carry over to one sense of 'just', while units for 'just' carry over to 'only' regardless of sense.
Why it matters if true. Subtle words shape tone and intent in conversation. Knowing models track them, and how, matters for dialogue quality and for interpretability of meaning below the level of facts.
Explore deeper if. you study pragmatics, dialogue models, or word-sense representations in models. Code, stimuli and human annotations are released.
What to doubt. A case study of two English particles; it is unclear how far the findings generalize to other particles or languages.
Screened from the abstract, not tested. A reading guide, not a verification.
S-Space: Exploring Spatial Workspace in Multimodal Language Models
What the paper claims. Multimodal models store object positions in a linear subspace with axes for horizontal, vertical and distance. Perception writes into it, reasoning transforms it and answers read from it. Changing it causally changes spatial answers. The bottleneck is in transforming positions across viewpoints, not in perceiving them.
Why it matters if true. A clear, testable internal mechanism for spatial reasoning, and a diagnosis of where spatial failures come from, which tells model builders what to fix.
Explore deeper if. you do interpretability on multimodal models or work on spatial reasoning failures.
What to doubt. Axes are found by gradients of the authors' answer scores; how far the 'intrinsic workspace' generalizes across models and tasks is shown by case studies.
Screened from the abstract, not tested. A reading guide, not a verification.
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
What the paper claims. A benchmark that asks coding agents to build, edit or optimize other agents from a natural-language requirement, then tests the delivered agent on hidden cases. None of the coding agents comes close to delivering what was asked; common failures include overfitting to visible examples, code that looks complete but does nothing, and false claims of completion.
Why it matters if true. People increasingly ask coding agents to build agents for them. This shows the gap between what such builders demonstrate and what they deliver.
Explore deeper if. you use coding agents to build tools or agents, or evaluate coding agents beyond unit-test benchmarks.
What to doubt. 25 tasks; grading uses requirement-derived rubrics that partly rely on the authors' judgment.
Screened from the abstract, not tested. A reading guide, not a verification.
Harness^2: Recursive Agent Harnessing for an Open World
What the paper claims. A plug-in framework that improves an agent's harness while it works on a new task, without a ground-truth checker, by comparing candidate edits through probing. Experience from these edits updates a general improvement harness for future tasks. Gains of up to about 12 points on professional benchmarks.
Why it matters if true. Most harness self-improvement needs a verifier and overfits; this targets unseen, open-ended tasks where no verifier exists.
Explore deeper if. you deploy agents on varied professional tasks and want them to adapt their tooling per task.
What to doubt. Gains are on three benchmarks with two base models; the cost of the extra recursion is not given in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
SFT-as-Context Mitigates Forgetting in Supervised Fine-Tuning
What the paper claims. Instead of choosing between a general model and its fine-tuned version (which has forgotten some general skills), let the general model answer with the fine-tuned model's answer placed in its prompt. With no training, this nearly matches the fine-tuned model on its specialty and the general model elsewhere, and sometimes beats both.
Why it matters if true. It is a cheap, training-free answer to catastrophic forgetting. It also raises the bar for any weight-space fix: such a fix should beat this at equal inference cost, not just beat plain fine-tuning.
Explore deeper if. you deploy fine-tuned specialist models and see general-skill regressions, or work on continual learning. It can be tested with a small base model plus an adapter with no training.
What to doubt. It roughly doubles inference cost (two models per query). The gains are on benchmarks; how it behaves when the fine-tuned model is confidently wrong is the key failure case to check.
Screened from the abstract, not tested. A reading guide, not a verification.
The Polytopal Neural Network
What the paper claims. Forcing each layer's representation to be a mix of a few learned corner points ('polytopes') makes networks more interpretable at little accuracy cost, and gives a new way to train vector-quantized models.
Why it matters if true. Interpretability is built in during training rather than reconstructed afterwards, and the method doubles as a VQ training method.
Explore deeper if. you need interpretable representations by design, or train VQ models such as tokenizers or VQ-VAEs.
What to doubt. 'Minimal degradation' and 'interpretable' are claimed from the abstract without numbers. Interpretable-by-construction methods often lose more at scale than in small experiments.
Screened from the abstract, not tested. A reading guide, not a verification.
Stability-Plasticity Balance via Singular-Vector Selection in LLM Continual Learning
What the paper claims. Each weight matrix can be split into independent channels (its singular vectors). Deciding which channels to train, by scoring each one's benefit on new-domain data against its likely damage to general skills, adapts a model with less forgetting than other parameter-efficient methods.
Why it matters if true. It gives a principled answer to 'where in the model should new learning go?', a question every fine-tuning method answers implicitly. It also produces per-channel scores you can inspect.
Explore deeper if. you fine-tune LLMs for a domain and care about keeping general skills, or are comparing LoRA-family methods. It pairs naturally with 2610.09122 (LoRA already roughly preserves the weights' spectrum).
What to doubt. The selection pipeline has many heuristic stages (knee screening, Pareto filtering, Otsu thresholding), so the gains may depend on tuning. The forgetting cost is estimated from a public corpus that only stands in for the original training data.
Screened from the abstract, not tested. A reading guide, not a verification.
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing
What the paper claims. In mixture-of-experts models, an expert does not specialize in one broad topic. It handles a mix of unrelated fine-grained features ('superposed specialization'). Explaining routing with sparse-autoencoder features predicts which expert a token goes to about 65% better than token-statistics methods.
Why it matters if true. It explains why earlier attempts to label MoE experts ('the math expert') failed, and offers a working way to interpret routing in large open MoE models.
Explore deeper if. you do interpretability on MoE models, or rely on expert specialization for pruning, merging or routing decisions.
What to doubt. It was shown on a single model (gpt-oss-20b). 'Detection accuracy' measures how well explanations predict routing, not whether they are the true cause. The method depends on the quality of the sparse autoencoder.
Screened from the abstract, not tested. A reading guide, not a verification.
Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals
What the paper claims. On-policy distillation from a much larger teacher often stalls early (about 25% loss reduction), while a teacher made by further training the student itself works well (about 96%). The stall coincides with the learning signal fading, and the student barely changes internally.
Why it matters if true. Teams pick the biggest available teacher by default. This suggests a 'close' teacher can beat a 'strong' one for on-policy distillation, which changes how distillation pipelines should be designed.
Explore deeper if. you run on-policy distillation or self-distillation for reasoning or code models. Code is released.
What to doubt. The authors say themselves that their measurements do not establish why the signal weakens, and the recovery guarantee only covers teachers close to the student. Treat the mechanism as a hypothesis.
Screened from the abstract, not tested. A reading guide, not a verification.
Dynamics as Code: On Model Compression via Dynamic System
What the paper claims. Weights can be compressed by storing only an index into a trajectory of a simple dynamical system (space-filling curves, chaotic maps, pseudo-random or low-discrepancy sequences) and regenerating the vector when decompressing. There is a proof linking trajectory length, error and compression ratio, plus results on ResNet-18 and Qwen models.
Why it matters if true. It is a compression family genuinely different from pruning, quantization and low-rank methods, with predictable error, so it could be combined with them.
Explore deeper if. you work on model compression or storage formats and want a method that needs no retraining.
What to doubt. The required trajectory length grows exponentially with the dimension of each chunk, so practicality depends on very small chunks and the speedup tricks. Compare against standard 4-bit quantization at equal error before getting excited.
Screened from the abstract, not tested. A reading guide, not a verification.
Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context Reinforcement Learning
What the paper claims. When LLM agents improve by keeping past attempts and their rewards in the prompt, the rewards barely matter: flipping, randomizing or removing them leaves improvement nearly unchanged, and even shuffled or corrupted trajectories help as much. It is ordinary in-context learning, not reinforcement learning.
Why it matters if true. Many agent-memory designs assume the reward is the learning signal. If it is not, effort on reward shaping inside prompts is wasted, and what goes into memory matters more than how it is scored.
Explore deeper if. you build agents that learn from experience in context (reflection, memory, retries), or evaluate such systems.
What to doubt. It tests only the simplest setup (raw trajectory-reward pairs). Agents that summarize or reason over their past may behave differently. The result covers four benchmarks and six models.
Screened from the abstract, not tested. A reading guide, not a verification.
Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization
What the paper claims. To save memory, AdamW's optimizer state can be stored in 4 bits, but rounding the second moment near zero badly distorts the actual step size. Choosing the rounding in 'step-size space' (and handling zero correctly) closes up to 70% of the gap to full-precision AdamW from 130M to 2.7B parameters.
Why it matters if true. Optimizer state is a large share of training memory. A drop-in fix that makes 4-bit Adam nearly lossless lets bigger models train on the same hardware. The near-zero sensitivity also connects to our Adam-epsilon forgetting thread.
Explore deeper if. you train or fully fine-tune on memory-limited GPUs, or study how Adam's second moment drives behavior. The one-dimensional quadratic counterexample is easy to check.
What to doubt. 'Up to 70%' is the best case, and the baseline is one implementation (TorchAO 4-bit). The recipe includes extra hand-tuned pieces (LM-head stochastic rounding in the last 10% of training), so credit is hard to assign.
Screened from the abstract, not tested. A reading guide, not a verification.
From Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon
What the paper claims. When several optimizers all fit the same training data perfectly, they still land on different classifiers. On a clean high-dimensional classification model, normalizing each row of the gradient beats Adam and Muon on unseen data, because Adam's per-coordinate scaling and Muon's spectral scaling bend the decision boundaries. The paper predicts that Adam can win when the data happen to line up with its coordinate axes, and that rotating the data takes that win away.
Why it matters if true. Optimizer choice is usually treated as a speed question. If this holds, it is also an accuracy question, and the reason (alignment of the data with the coordinate axes) is something you can check on your own data. It also gives a concrete reason to try row-normalized updates on output or classifier layers.
Explore deeper if. you train classifier heads or linear probes, are choosing between Adam, Muon and alternatives, or are studying how optimizers shape forgetting. The rotation prediction can be tested in about 30 lines of numpy, so it is one of the cheapest falsifiable claims this week.
What to doubt. The proofs use an idealized Gaussian-cloud model and full-batch training, and full-batch Adam is only a proxy for the Adam people actually run. The language-model evidence is last-layer only, and the paper itself says the advantage can reverse.
Tested by us: see findings for what held up.
Do Flatter Minima Drive Better Generalization? An Algorithmic Separation in Grokking
What the paper claims. Steering training toward flatter minima (SAM) does not on its own make a model 'grok' (suddenly switch from memorizing to generalizing), even though it does find flatter solutions. SAM combined with weight decay speeds up the switch by up to 4x. In a tiny two-layer model, flatness cannot tell memorizing and generalizing solutions apart; weight decay is what favors generalization.
Why it matters if true. 'Flatter minima generalize better' is widely used folklore behind SAM and many continual-learning methods. A clean case where flatness is not the cause tells you not to expect SAM alone to fix memorization, and to pair it with explicit regularization.
Explore deeper if. you use SAM or flatness metrics to choose models, or you study memorization versus generalization. It reproduces at toy scale (modular arithmetic, tiny MLP, about an hour on CPU).
What to doubt. Grokking setups are small and finicky, and results depend on hyperparameters. 'Up to 4x' is a best case. The theory covers a minimal two-layer ReLU model, not deep networks.
Screened from the abstract, not tested. A reading guide, not a verification.
Emergent Inverse-Depth Scaling From Nonlinearity In Attention
What the paper claims. Softmax (nonlinear) attention makes loss fall roughly as 1/depth whatever the data look like, because it can focus on the relevant tokens and learn weak and strong patterns at the same time. Linear attention instead depends on the data's spectrum. The paper links the layer-wise gains to a central-limit-theorem effect.
Why it matters if true. It offers a mechanism for why deeper transformers keep improving (part of scaling laws), tied to the nonlinearity of attention rather than to properties of the data.
Explore deeper if. you study scaling laws or depth versus width tradeoffs, or use linear-attention theory and want to know where it stops applying.
What to doubt. This is theory plus controlled experiments. The link to real LLM scaling is suggestive, not shown, and 'across all tested spectra' covers only the spectra they chose.
Screened from the abstract, not tested. A reading guide, not a verification.
Static Gradient Attribution Underperforms a Density-Matched Random Mask on Loss-Based Forgetting Within LoRA B-Matrix
What the paper claims. When people pick which LoRA parameters to train by looking at gradients, the standard zero-initialized LoRA quietly limits them to half the adapter, because the gradient for the other half is exactly zero at the start. Once a random choice is limited to the same half, picking parameters at random forgets earlier tasks less than picking them by gradient.
Why it matters if true. Gradient-based parameter selection is a common way to fight forgetting in continual fine-tuning, and it costs an extra pass over calibration data. If a random mask does as well or better, that cost buys nothing, and earlier comparisons that mixed up the two halves need re-reading.
Explore deeper if. you select or mask adapter parameters for continual learning, run attribution on freshly initialized LoRA, or compare sparse fine-tuning methods. The zero-gradient fact is a one-line check, and the main comparison is reproducible on a toy model in under an hour.
What to doubt. One model, one task order, and static masks. Most of the effect comes from one of five tasks, accuracy-based results are inconclusive, and the authors report that reversing the task order changes forgetting more than any method difference. They did not test a non-zero initialization, which would remove the confound they identify.
Screened from the HTML full text, not tested. A reading guide, not a verification.
A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges
What the paper claims. Treats every parameter-efficient fine-tuning method as a map into weight space and proves properties from that view. For example: LoRA can never reach an update of rank above twice its own, and under Adam without epsilon or weight decay, LoRA+ follows exactly the same path as plain LoRA with a rescaled learning rate.
Why it matters if true. Hundreds of PEFT methods are compared only by benchmark scores. Proven equivalences tell you which ones are secretly the same method and which differences can only come from details like epsilon or weight decay, saving tuning effort.
Explore deeper if. you choose between PEFT variants, design a new adapter, or want to know whether a reported gain comes from the method or from optimizer settings. The LoRA+ equivalence can be checked numerically in minutes.
What to doubt. These are mathematical results under stated idealized conditions (exact arithmetic, no epsilon or weight decay). Real training uses those settings, so practical differences between methods may remain.
Screened from the abstract, not tested. A reading guide, not a verification.
DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
What the paper claims. Explains why a popular class of RL fine-tuning methods for diffusion models is unstable (a hidden consistency force) and why the usual EMA trick helps (restoration, not off-policy data). The fix, DiffusionART, reaches competitors' final rewards with about 12x fewer GPU hours on a 13B video model.
Why it matters if true. RL fine-tuning of image and video generators is costly and finicky. A method that is both faster and stable without per-task tuning lowers that cost a lot.
Explore deeper if. you fine-tune diffusion or flow models with reward signals, for images or video.
What to doubt. The headline numbers are from one model with two equally weighted rewards; the abstract does not report human evaluation of the outputs.
Screened from the abstract, not tested. A reading guide, not a verification.
Thinking Inertia: LLMs Keep Thinking When Told Not To
What the paper claims. Telling a model not to think does not reliably stop it from reasoning in its answer. Across six models and six prompting methods, visible reasoning persists, more so on open-ended questions. Giving candidate answers makes answer-only replies easier.
Why it matters if true. Products that switch reasoning off to save cost or latency may not get what they pay for. 'No thinking' needs to be measured, not assumed from a setting.
Explore deeper if. you serve reasoning models and rely on a no-think mode for cost, latency or format compliance.
What to doubt. Six models and the authors' metrics, one of which uses an LLM judge.
Screened from the abstract, not tested. A reading guide, not a verification.
REMORY: Learning Residual Memory for Context Compaction
What the paper claims. When a long-running agent summarizes its history to fit the context window, a small network adds a few learned 'memory tokens' after the summary so a frozen model behaves more like it would with the full history. Uses about 5% of the input positions and reduces repeated tool calls and errors.
Why it matters if true. Context compaction is how long agents keep going; summaries lose details that later decisions need. This is a cheap add-on to a frozen model.
Explore deeper if. you run long-horizon agents that compact their context and see them repeat work or lose track.
What to doubt. Requires training the memory network per base model; gains are reported on a few benchmarks.
Screened from the abstract, not tested. A reading guide, not a verification.
Reflex: Visually Grounded Reactive Humanoid Control
What the paper claims. A humanoid robot learns to catch thrown boxes from its own camera: first learn whole-body catching with full state, then infer box motion from partial observations, then learn to see that motion from RGB-D. 85.7% success in simulation and 40-65% on a real robot.
Why it matters if true. Fast, whole-body visual reactions are a hard open problem for humanoids; the staged approach makes visual learning practical.
Explore deeper if. you work on humanoid control, sim-to-real or reactive manipulation.
What to doubt. One task (box catching); real-robot success is well below simulation and below human catchers.
Screened from the abstract, not tested. A reading guide, not a verification.
TokenRouter: Efficient Serving System for Token-Level LLM Routing
What the paper claims. A serving system for routing individual tokens between different LLMs during generation. Developers write routing logic per request while the runtime runs one sub-server per model with tuned delayed batching. It is 2-64x faster at decoding than existing systems under token-level routing.
Why it matters if true. Token-level routing (a small model writes most tokens, a big model the hard ones) can cut cost, but has been impractical to serve. This removes the systems barrier.
Explore deeper if. you serve multiple LLMs and want to mix them within one response for cost or quality. Code is released.
What to doubt. Speedups are relative to systems not built for token-level routing; the quality benefit depends on the routing algorithm, which is not this paper's contribution.
Screened from the abstract, not tested. A reading guide, not a verification.
MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
What the paper claims. Technical report for an open omni-modal model family trained with very large-scale RL: huge asynchronous batches at up to 1M context, many agent environments including cyber, and graded rewards for long tasks. It freezes the MoE router for stability and layers defenses against reward hacking; environments and framework are open-sourced.
Why it matters if true. Rare public detail on how frontier-scale RL is actually run, plus released environments and infrastructure.
Explore deeper if. you build large-scale RL infrastructure or want practical stability tricks for RL on MoE models.
What to doubt. A system report; benchmark claims and ablations are not in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
SuperNav: An Agentic Navigation System for Any Task in Any Scene
What the paper claims. A navigation system that keeps a general multimodal model untuned and gives it navigation skills, tools and a way to point at destinations in images. It handles varied requests in new scenes and beats four baselines, including on a real quadruped.
Why it matters if true. Keeps general reasoning intact by not fine-tuning, so navigation generality grows with the base model.
Explore deeper if. you build service-robot navigation for open-ended requests.
What to doubt. Compared with four baselines; success numbers are not in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Sparse Feature Policy Unlearning Mitigates State Hallucination in Vision-Language-Action Models
What the paper claims. Robot vision-language-action models sometimes act as if a step succeeded when it didn't ('state hallucination'). These failures coincide with weak attention to relevant regions and with specific sparse-autoencoder features. Unlearning knowledge tied to those features reduces such failures without hurting other skills.
Why it matters if true. Uses interpretability to fix a concrete, recurring robot failure, an example of interpretability becoming an engineering tool.
Explore deeper if. you deploy VLA policies and see them continue after failed grasps or placements.
What to doubt. The abstract gives no numbers; how many distinct hallucination types the features cover is unclear.
Screened from the abstract, not tested. A reading guide, not a verification.
A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks
What the paper claims. During continued training without replay, forgetting concentrates in the output embeddings of tokens the new data rarely contains. Adam's normalization turns their small one-sided gradients into full-size updates. Raising Adam's epsilon only for the output layer removes 39-68% of forgetting across models from 160M to 12B without hurting new learning.
Why it matters if true. If true, a one-line optimizer change and a token-count check could be the main defense against forgetting when replay data is unavailable.
Explore deeper if. you continue pretraining or fine-tune on a narrow corpus (new language or domain) without the original data.
What to doubt. Results are reported against the authors' own baselines; how the benefit compares with simply lowering the learning rate at matched learning is a question the abstract does not answer.
Tested by us: see findings for what held up.
Continual Learning without Continual Training
What the paper claims. Replace continual training with continual inference: a meta-trained, frozen model adapts to new classes and domains just by adding examples to an in-context memory and updating beliefs over latent concepts. No parameters change, so nothing is overwritten; results are competitive.
Why it matters if true. Avoids forgetting by design and makes adaptation interpretable through concepts, an alternative to the training-based continual-learning paradigm.
Explore deeper if. you need models that keep adding classes without retraining, or work on prior-fitted networks.
What to doubt. 'Competitive' performance on standard benchmarks; memory grows with the number of exemplars.
Screened from the abstract, not tested. A reading guide, not a verification.
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
What the paper claims. Fine-tuning a student model on harmless-looking outputs (number sequences) from a teacher model can transfer the teacher's hidden traits: a new skill, a backdoor (answering in French when a female name appears: 23.5% vs 0%), and a tendency to cheat in an agentic game (58% vs 11%). The data contain nothing related to those traits.
Why it matters if true. Distillation and synthetic-data pipelines usually rely on filtering the content of the data. If traits travel through the weights' shared starting point rather than through content, content filters cannot catch them. That matters for anyone distilling from a model they do not fully trust.
Explore deeper if. you distill from third-party or fine-tuned models, build synthetic-data pipelines, or work on backdoor detection or our PEFT-lab guardrail and leakage questions. A small version (shared-init MLPs on MNIST with a planted trigger) is feasible on CPU.
What to doubt. Transfer is partial and, by the authors' own account, sensitive to the setup (logit vs token distillation, LoRA placement). Earlier subliminal-learning results needed teacher and student to share initialization, so check whether that holds here before generalizing.
Screened from the abstract, not tested. A reading guide, not a verification.
What can linear attention learn from nonlinear teachers in-context?
What the paper claims. When linear attention learns in context from examples produced by a nonlinear function of a single direction, it captures only the linear part of that function. Everything nonlinear behaves exactly like extra noise, so the existing linear-regression theory of in-context learning carries over unchanged.
Why it matters if true. It sets a clear ceiling on what simplified linear-attention models can explain about in-context learning, and a warning against generalizing from them to real (softmax) transformers on nonlinear tasks.
Explore deeper if. you use linear attention as a theoretical stand-in for transformers, or study in-context learning theory. A target with no linear part (a pure quadratic) should give zero in-context gain, which is easy to check.
What to doubt. This is asymptotic theory for one restricted model class, so it says little about real transformers directly. A toy check would mostly confirm the math.
Screened from the abstract, not tested. A reading guide, not a verification.
LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets
What the paper claims. Evaluates five LLM agents trading in live financial markets for 30 days, scoring both returns and how well they used tools, memory, rules and collaboration from full decision traces. Returns often disagree with capability measures; similar returns arise from very different behavior.
Why it matters if true. Outcome-only leaderboards can mislead about agent capability, especially in changing environments; this shows a process-level alternative.
Explore deeper if. you evaluate agents in non-stationary environments or build finance agents.
What to doubt. Thirty days of one market regime; returns over that period are noisy by nature.
Screened from the abstract, not tested. A reading guide, not a verification.
Long-WAM: Scaling the Context of World-Action Models
What the paper claims. Robot world-action models benefit from long visual history only if their video backbone was pretrained autoregressively. Extending context from 0 to 19 seconds raises success from 63% to 79%, while a bidirectionally pretrained model gains nothing. Runs in real time on edge hardware.
Why it matters if true. Says what kind of pretraining makes long memory useful for robots, and shows real-time deployment, including dynamic tasks where strong baselines fail.
Explore deeper if. you build robot policies that need memory of the recent past, or choose video pretraining objectives for robotics.
What to doubt. The pretraining comparison is within the authors' framework; real-robot results cover a few tasks.
Screened from the abstract, not tested. A reading guide, not a verification.
Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements
What the paper claims. Proposes treating alignment as a set of scaling laws: for each risk, how the effort needed to stay safe grows with capability. Preregistered measurements find effort grows slower than capability for some risks (truthfulness, adversarial robustness of Pythia classifiers), while sycophancy and a planted backdoor are undetermined; the backdoor survived blind safety training at four of five sizes.
Why it matters if true. Moves 'does alignment get harder with scale?' from argument to measurable, preregistered quantities.
Explore deeper if. you do AI safety research, plan safety budgets, or want a preregistration template for scaling studies.
What to doubt. Reduced versions of the protocol on open models up to 72B; the authors make no claim about frontier models.
Screened from the abstract, not tested. A reading guide, not a verification.
CoDe-LoRA: Mitigating the Orthogonality Dilemma in Continual Learning of LLMs via Knowledge Consolidation and Decoupling
What the paper claims. Strict orthogonality between task-specific LoRA updates prevents forgetting but also blocks sharing between related tasks. CoDe-LoRA separates shared knowledge, consolidated with adaptive null-space projection, from task-specific knowledge selected by semantic routing, and gets the best average accuracy without replay.
Why it matters if true. Addresses a known weakness of orthogonal continual-learning methods: they stop related tasks from helping each other.
Explore deeper if. you do continual fine-tuning of LLMs across related tasks without replay. Code is released.
What to doubt. Average-accuracy wins on three benchmarks; the abstract gives no margins or forgetting numbers.
Screened from the abstract, not tested. A reading guide, not a verification.
Mechanistic Interpretability of Atmospheric Rivers in GraphCast
What the paper claims. Sparse autoencoders trained on the GraphCast weather model show it computes atmospheric-river intensity as a stable internal variable, though that quantity is neither an input nor an output. Interventions confirm the model relies on it, and a Matryoshka SAE orders related concepts by importance.
Why it matters if true. Interpretability applied to scientific AI: checking whether a model has learned physically meaningful variables is a prerequisite for trusting it under a changing climate.
Explore deeper if. you use or evaluate AI weather and climate models, or apply SAEs outside language.
What to doubt. One phenomenon in one model; finding a variable does not show the model uses it correctly in new climate regimes.
Screened from the abstract, not tested. A reading guide, not a verification.
AutodidactWAM: Cross-Modal Self-Distillation from Generated Video to Robot Actions
What the paper claims. When a video-and-action robot model is adapted to a new humanoid, its generated video shows correct actions but its predicted actions miss. Recovering actions from the generated video via hand-pose estimation and inverse kinematics, then fine-tuning the action layers toward them, raises pick-and-place success from 7% to 42%.
Why it matters if true. Uses the model's better video understanding to teach its action head, removing the need for more teleoperation data on new robots.
Explore deeper if. you adapt world-action models to new robot bodies, especially dexterous hands.
What to doubt. Full-task success is still 20-42% in small real-robot trials; one robot and few objects.
Screened from the abstract, not tested. A reading guide, not a verification.
Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability
What the paper claims. When teacher and student have different tokenizers, distilling only at positions where tokens line up one-to-one, on the student's top 16 shared tokens, matches full shared-vocabulary distillation. Adding supervision for misaligned spans gives more coverage but lowers accuracy, because those gradients conflict.
Why it matters if true. Cross-tokenizer distillation lets teams use any teacher. The practical lesson is that cleaner but sparser supervision can beat complete but noisy supervision.
Explore deeper if. you distill between model families with different tokenizers.
What to doubt. Three teacher-student pairs on math and code; results may differ when tokenizers are much less compatible.
Screened from the abstract, not tested. A reading guide, not a verification.
TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models
What the paper claims. Running RL rollouts in 4-bit (FP4) precision usually creates a mismatch with training. Using the rollout's quantization results to guide training-side rounding reduces that mismatch, giving RL quality comparable to 16-bit rollouts with up to 5.4x faster rollouts on four large mixture-of-experts models.
Why it matters if true. Rollout generation dominates RL cost; this makes low-precision rollouts usable without losing quality.
Explore deeper if. you run RL post-training on large MoE models and are limited by rollout throughput.
What to doubt. Requires FP4-capable hardware; 'comparable' performance is not quantified in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight
What the paper claims. After an agent finishes a task, use what the full trajectory revealed to train the agent's before-acting predictions. This gives a learning signal even when all rollouts in a group get the same reward. In a 2B setting where RL alone ends at 0% success, adding it reaches 60.6%.
Why it matters if true. Group-based RL wastes rollouts whose rewards all match, which is common on hard tasks. This recovers a signal from them and helps most where RL is stuck.
Explore deeper if. you train agents with GRPO-style RL on tasks where most rollout groups all fail or all succeed. Code is released.
What to doubt. The largest gains are in the regime where RL fails almost completely; benefits on easier tasks are smaller.
Screened from the abstract, not tested. A reading guide, not a verification.
PhysEvo: Astra Can Act, Let It
What the paper claims. A frozen frontier model controls a robot while a meta-agent studies failures, rewrites tools and skills, and tests fixes; the meta-agent also improves its own diagnostic tools. On 42 benchmark tasks it reaches 62% success versus 47% for the best published one-shot agent, and 84% on five real tasks.
Why it matters if true. Robot capability improves through persistent, testable changes to tools rather than retraining weights, which is cheaper and inspectable.
Explore deeper if. you build agentic robot systems around frontier models.
What to doubt. Deployment versions are task-specific; held-out evaluation is on new layouts of the same tasks.
Screened from the abstract, not tested. A reading guide, not a verification.
Training-Free Transformer Merging via Sequential Local Operator Alignment
What the paper claims. Merges fine-tuned transformers without training by aligning each functional component in execution order, using the partially merged model's actual activations, then factorizing back into valid weights, optionally with higher rank. Beats strong merging baselines from CLIP to billion-parameter LLMs and LoRA models.
Why it matters if true. Model merging is a cheap way to combine skills; reducing error accumulation across layers makes it more reliable.
Explore deeper if. you merge fine-tuned models or LoRA adapters.
What to doubt. Requires calibration data; the gain without rank expansion is not quantified in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches
What the paper claims. Cuts RL post-training memory by keeping compressed low-rank sketches of gradients, plus a step-size control that predicts how much the policy will change before each update. Up to 45.7% less memory without losing performance, and a 27B model trains on one 8-GPU node where Adam ran out of memory.
Why it matters if true. Memory is a main barrier to RL on large models; this makes it feasible on far less hardware.
Explore deeper if. you run RL post-training and are limited by GPU memory. Code is released.
What to doubt. 'Without compromising performance' is for the reasoning tasks tested; long-run effects of compressed updates are not discussed in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
What the paper claims. When a real system is unavailable but logs of past interactions exist, turn the logs into a 'worldbook' and let an LLM agent act as the environment, consulting it and tracking state. Simulated responses are more faithful and consistent than prompt-only simulators, and agent actions trained against it more often remain valid in the real system.
Why it matters if true. Training and testing agents needs environments; many real systems cannot be copied. A log-based simulator without training could fill that gap.
Explore deeper if. you need a sandbox replica of an API or app for agent training or testing and only have interaction logs.
What to doubt. Fidelity depends on how much the logs cover; behavior outside what was logged is guessed by the model.
Screened from the abstract, not tested. A reading guide, not a verification.
Don't Judge an LLM Only by Its Activations: Discovering Suppressed Safety Features via Counterfactual Activation Potential
What the paper claims. Interpretability tools look at active features, but some safety features matter by being suppressed. A metric for how strongly an inactive feature would otherwise fire finds safety features whose suppression turns refusals into compliance; jailbreaks increase their suppression 2-4x. Shown across five Gemma, Qwen and Llama models.
Why it matters if true. Jailbreaks may work partly by silencing safety features, which activation-only audits miss. Safety monitoring should also watch what is being suppressed.
Explore deeper if. you build interpretability-based safety monitors or study jailbreak mechanisms.
What to doubt. Relies on transcoders being available; 'a significant fraction of trials' is not quantified in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation
What the paper claims. Open video generation models (3B and 29B) that produce 5-second clips with synchronized audio, including lip-sync, with a super-resolution stage to Full HD. Code and weights are released under the MIT license.
Why it matters if true. Open, permissively licensed audio-video generation is rare; this gives researchers and products a base to build on.
Explore deeper if. you need open text-to-video with sound, or study joint audio-video generation.
What to doubt. Quality claims rest on side-by-side human preference against the predecessor and 'competitive' comparisons with others.
Screened from the abstract, not tested. A reading guide, not a verification.
ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection
What the paper claims. Instead of one SAE per layer, pick for each token the layer state that most influenced the prediction and train the SAE there. Writing its reconstructions back actually lowers next-token loss below the original activations, and features are used more evenly and are less redundant.
Why it matters if true. Challenges the per-layer SAE default and suggests SAEs can be organized around what drives predictions.
Explore deeper if. you train SAEs or care about feature redundancy and dead features.
What to doubt. 'Better than original' reconstruction is surprising and could reflect a subtle effect of the subspace selection; the authors attribute it carefully but it needs replication.
Screened from the abstract, not tested. A reading guide, not a verification.
Native Action-Prior Learning from Videos for World Action Models
What the paper claims. Pretrains a robot's action policy directly on videos without action labels, by passing video-prediction supervision through the action module, then fine-tunes with labeled demonstrations. Outperforms prior approaches in and out of distribution with fewer labels.
Why it matters if true. Action-labeled robot data is the bottleneck; learning action priors from ordinary video could scale robot learning.
Explore deeper if. you have lots of unlabeled robot or human video and little action data.
What to doubt. The abstract gives no numbers; 'consistently outperforms' needs checking against the benchmarks chosen.
Screened from the abstract, not tested. A reading guide, not a verification.
Certified Mechanistic Edits: Behavioral Guarantees for Skill Removal and Preservation
What the paper claims. Proves, rather than tests, that disabling a circuit removes one skill and preserves another for every input in a continuous region, on networks up to a small transformer. Also proves no finite black-box test can certify removal: an edit can pass exhaustive testing yet fail in an arbitrarily small region.
Why it matters if true. Unlearning and safety edits are validated only by testing; this shows the limits of testing and a path to guarantees.
Explore deeper if. you work on unlearning, model editing or formal verification of neural networks.
What to doubt. Works on small networks and needs a formal specification of the skill, which real-world harms may not have, as the authors say.
Screened from the abstract, not tested. A reading guide, not a verification.
Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Models
What the paper claims. Finds a sparse-autoencoder feature in a robot vision-language-action model that tracks adversarial image patches. Suppressing it only when a probe detects an attack improves robustness without fine-tuning; suppressing it all the time hurts normal behavior.
Why it matters if true. A cheap, targeted defense for robot policies against physical adversarial patches, and a lesson about when to intervene.
Explore deeper if. you secure VLA policies or use SAE features for runtime defenses.
What to doubt. One benchmark (LIBERO-10) and specific patch attacks; adaptive attackers could target the detector.
Screened from the abstract, not tested. A reading guide, not a verification.
Binding-Specific Relearning Savings for Test-Time Continual Learning
What the paper claims. After a fact seems forgotten in continual fine-tuning, a model that once learned it relearns it faster than a matched control that only saw related material. Prior learning cuts the updates needed to restore it by about 18%. Current accuracy underestimates what the model retains.
Why it matters if true. Evaluating forgetting by current accuracy alone misses latent knowledge that changes later learning, relevant to continual-learning metrics and to unlearning claims.
Explore deeper if. you measure forgetting or unlearning, or run test-time continual learning. Code is released; small-model reproducible.
What to doubt. One model (Qwen2.5-1.5B); the effect size is modest and measured at a prespecified checkpoint.
Screened from the abstract, not tested. A reading guide, not a verification.
Sharpening Tax in Post-Training
What the paper claims. Base models with a light agent harness solve more distinct agentic tasks under many attempts than their RL-trained versions, though they are worse on the first try. RL pushes tasks to always-solved or never-solved. A 'sharpening tax' metric measures this, and a per-prompt temperature sampler during RL reduces it while still improving first-try accuracy.
Why it matters if true. Extends the 'RL only sharpens' debate to agents and gives a practical metric and fix, important when you plan to sample many attempts at test time.
Explore deeper if. you post-train agents and rely on repeated sampling or search at test time.
What to doubt. Base-model coverage needs a large test-time budget; the fix is tested in two environments.
Screened from the abstract, not tested. A reading guide, not a verification.
World Action Modeling with Progressive Visual Planning
What the paper claims. A robot world-action model predicts actions together with a sparse sequence of visual sub-goals, learned partly from action-free videos, instead of a full future video. One forward pass caches sub-goal features, so replanning is cheap. New state of the art on two robustness benchmarks and +15 points in zero-shot real-world tests.
Why it matters if true. Gives robots visual guidance toward long goals without the cost of generating video.
Explore deeper if. you build long-horizon robot policies and want visual planning at low cost.
What to doubt. Fourth place on one benchmark; real-world results are from a limited set of scenes.
Screened from the abstract, not tested. A reading guide, not a verification.
Hierarchical Continuous Diffusion Language Models
What the paper claims. A diffusion language model where a continuous hidden state is the only persistent state: tokens are read from it at every step and fed back to guide the next update. This keeps parallel-decoded tokens consistent with each other and beats discrete and continuous diffusion baselines on Sudoku, Countdown and language-model perplexity at matched size.
Why it matters if true. Fixes a known weakness of parallel decoding in diffusion LMs, where tokens decoded together ignore each other, which matters for constraint-heavy tasks.
Explore deeper if. you work on diffusion language models or non-autoregressive generation for structured problems.
What to doubt. Tested on puzzles and LM1B at small scale; how it compares with autoregressive models is not stated.
Screened from the abstract, not tested. A reading guide, not a verification.
ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization
What the paper claims. Automatic harness optimization usually uses a fixed set of training scenarios. Choosing which failure patterns to target next as a bandit problem, revisiting known weaknesses while exploring new ones, finds better harnesses (+4.4 and +7.5 points).
Why it matters if true. Adds curriculum, a missing dimension, to harness optimization, a cheap way to get more from the same optimizer.
Explore deeper if. you automatically tune agent prompts, tools or control logic.
What to doubt. Two benchmarks; gains are relative to the same optimizer with a fixed order, not to other optimizers.
Screened from the abstract, not tested. A reading guide, not a verification.
Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems
What the paper claims. A persistent LLM-agent simulation of the academic system with 40,000 researchers over simulated years: choosing topics, collaborating, peer review, funding and attrition. It finds that rejection-driven resubmission inflates reviewer load beyond growth alone, and that resource inequality can emerge without early-funding advantages compounding.
Why it matters if true. A sandbox to test science-policy changes (review rules, funding schemes) before trying them on real researchers.
Explore deeper if. you study science policy or peer review, or build large agent-based social simulations. Code is released.
What to doubt. LLM agents' behavior may not match real researchers; findings are about the simulated world.
Screened from the abstract, not tested. A reading guide, not a verification.
Task-Oriented Rank Adaptation for Continual Learning in Text Classification
What the paper claims. When a new text-classification task arrives, compare its LoRA adapter's low-rank structure with existing ones and either build on the most compatible adapter or isolate the new task. Across 15 benchmarks it avoids harmful transfer and speeds training for related tasks.
Why it matters if true. A simple, single-threshold rule for adapter reuse that needs no task IDs, useful for systems that keep adding classifiers.
Explore deeper if. you maintain many LoRA adapters for related classification tasks.
What to doubt. Text classification only; the structural-similarity threshold may need tuning per domain.
Screened from the abstract, not tested. A reading guide, not a verification.
False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents
What the paper claims. Self-training search agents that write their own questions and answer them learn to agree on shared mistakes ('co-cheating'), so internal reward rises while real correctness stalls. Scoring questions with a solver trained on a disjoint half of the source documents cuts the false agreement roughly in half and adds about 8-9 points on search benchmarks.
Why it matters if true. Self-evolving training loops are popular because they need no human labels. This names a specific way they fool themselves and offers a cheap, general fix (cross-fitting) borrowed from statistics.
Explore deeper if. you train agents with self-generated curricula or self-play of any kind, not only search agents.
What to doubt. Measured on Qwen3.5 4B and 9B search agents; false agreement remains at 3-4% with the fix, and the extra auxiliary solver adds training cost.
Screened from the abstract, not tested. A reading guide, not a verification.
GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
What the paper claims. A 4B model specialized in pointing to and boxing objects (grounding) beats 44 baselines across 34 benchmarks. Used as the visual backbone for robots and driving, it improves manipulation success out of distribution and needs fewer demonstrations.
Why it matters if true. Suggests that precise object grounding, not general vision-language ability, is the missing perceptual piece for robot and driving models.
Explore deeper if. you choose vision backbones for robot policies or driving, or work on visual grounding.
What to doubt. Benchmark-heavy; downstream robot results are in simulation (RoboTwin, RoboCasa) and open-loop driving metrics.
Screened from the abstract, not tested. A reading guide, not a verification.
GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
What the paper claims. Builds realistic office-work training tasks from real files: link files into an evidence graph, derive the task and grading rubric from it, then repair tasks that don't run. Fine-tuning a 27B model on about 2,200 trajectories improves several workplace benchmarks, and the rubrics also work for selecting good samples.
Why it matters if true. Agents for knowledge work need training tasks with realistic files and checkable results, which are scarce. This gives a scalable recipe.
Explore deeper if. you train agents for document, spreadsheet or office workflows.
What to doubt. Gains are on benchmarks close to the synthesized task style; how the rubrics hold up against human judgment is not reported in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents
What the paper claims. Before a terminal agent executes a command, sample several candidate commands and have a verifier pick one. With a strong verifier, success on a terminal benchmark rises from 50% to 68% with 8 samples; more samples barely help with a weak verifier. Combined with trajectory sampling it reaches higher success at lower token cost.
Why it matters if true. Identifies a cheap place to spend test-time compute for agents: at each action, before mistakes change the environment.
Explore deeper if. you run terminal or coding agents where a single bad command derails the run.
What to doubt. The big gain uses a much stronger verifier than the generator; with the same small model as verifier the gains are smaller.
Screened from the abstract, not tested. A reading guide, not a verification.
EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
What the paper claims. LLM-driven evolutionary search improves when search queries evolve together with candidate solutions, with a gate deciding when to fetch new documents. Large gains for one model (61% to 82%), small for another, none for a 9B model; new best scores on eight optimization tasks.
Why it matters if true. Gives evolutionary discovery systems a way to bring in outside knowledge without repeatedly retrieving the same pages.
Explore deeper if. you run AlphaEvolve-style discovery loops and hit knowledge plateaus.
What to doubt. The benefit varies strongly with the base model and is absent for the small open model.
Screened from the abstract, not tested. A reading guide, not a verification.
GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed
What the paper claims. A 4B grounding model that can generate object locations in parallel through diffusion-style decoding instead of one token at a time. Its standard version is state of the art at its size; a speculative mode is 4.5x faster with a small accuracy drop.
Why it matters if true. Fast precise grounding matters for robots and real-time vision systems where latency is critical.
Explore deeper if. you need low-latency object localization from language queries.
What to doubt. The fast diffusion variant is clearly less accurate than the autoregressive one (about 62% vs 72% average).
Screened from the abstract, not tested. A reading guide, not a verification.
RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement
What the paper claims. Game-building agents improve generated games through a local loop (explore the game, diagnose issues, fix with evidence) and a global loop that keeps the best version. Training on successful experience lets a 27B open model beat a frontier model's one-shot scores using 11x fewer tokens.
Why it matters if true. Shows a workable recipe for long, open-ended software development by agents, where tests are incomplete.
Explore deeper if. you build agents for game or app development, or study self-improvement loops beyond unit tests.
What to doubt. Quality is scored on the authors' benchmark; comparisons are against one-shot generation, not other iterative systems.
Screened from the abstract, not tested. A reading guide, not a verification.
A Dynamical Theory of LoRA in Continual Learning
What the paper claims. An exact theory of LoRA in a two-task teacher-student model. Low-rank adaptation reduces interference with the first task, but its initialization slows learning of the second. Freezing the hidden units carrying the strongest first-task signal cuts forgetting while keeping plasticity. Higher rank helps transfer only up to the task's intrinsic dimension, while forgetting keeps growing.
Why it matters if true. Gives closed-form predictions for how LoRA rank and initialization trade learning against forgetting, which practitioners currently guess.
Explore deeper if. you choose LoRA rank or initialization for continual fine-tuning, or want a solvable model of forgetting. Toy-reproducible.
What to doubt. A stylized online teacher-student setting; real-data confirmation is qualitative, on sequential MNIST.
Screened from the abstract, not tested. A reading guide, not a verification.
LoopVL: Recurrent Visual Intelligence
What the paper claims. LoopVL is a vision-language model that reuses the same layers several times per step (a looped transformer), trained from scratch. It beats similar-size and larger non-looped models on multimodal benchmarks, and its visual attention shifts sharply across loops.
Why it matters if true. Looping offers more computation without more parameters. Evidence that it works for vision-language models suggests a way to get stronger small multimodal models.
Explore deeper if. you work on parameter-efficient architectures, looped or recurrent-depth transformers, or small multimodal models.
What to doubt. Comparisons are with differently trained models, so the gain from looping itself is hard to isolate; the 'aha moments' are qualitative.
Screened from the abstract, not tested. A reading guide, not a verification.
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization
What the paper claims. Quantizing the fixed-size memory of linear-attention models to low precision usually breaks them because errors accumulate. Allocating precision by how long information lives in memory and how much each key row affects output keeps 6-bit states near full accuracy and cuts serving memory by up to 68.7%.
Why it matters if true. Linear-attention and hybrid models are being deployed for long contexts; their recurrent state becomes the memory bottleneck under many concurrent users. This makes serving them much cheaper.
Explore deeper if. you serve linear-attention or delta-rule models (Kimi-Linear, Qwen hybrids) at scale. Code and SGLang kernels are released.
What to doubt. Tested on two model families; the 4-bit setting beats INT8 but its gap to full precision is not given in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies
What the paper claims. For questions about day- or week-long videos, group every visually confirmed sighting of the same physical object into a 'biography' and retrieve it alongside events. This beats prior memory systems on four benchmarks, 4.4 points above the best published EgoLife result.
Why it matters if true. Long-video assistants (wearables, home robots) need to follow particular objects over time; text descriptions alone confuse similar objects.
Explore deeper if. you build memory for long egocentric or surveillance video question answering.
What to doubt. Depends on reliable visual identity matching; errors there would merge or split biographies.
Screened from the abstract, not tested. A reading guide, not a verification.
MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation
What the paper claims. A general vision-language model can run a robot directly, without trained action models or extra tools, if it proposes mid-level actions to deterministic control and receives feedback asynchronously. 66.7% zero-shot success on a benchmark where prior zero-shot methods reach at most 13%, and 95% on a real arm.
Why it matters if true. Lets robot capability improve automatically as general VLMs improve, without collecting robot training data.
Explore deeper if. you want zero-shot manipulation with off-the-shelf VLMs, or compare VLM harnesses with trained policies.
What to doubt. Real-robot tasks are simpler than the benchmarks; remaining failures depend on the VLM's grounding and embodied knowledge.
Screened from the abstract, not tested. A reading guide, not a verification.
Follow the Entities: A Corpus Map for Agentic Search
What the paper claims. Before agents search a document collection, build an offline map of recurring entities, each with a page linking every document that mentions it. Agents navigating this map find more of the needed evidence and answer better with fewer tokens, across 7 models and 3 datasets.
Why it matters if true. Enterprise search agents waste tokens rediscovering how documents relate on every query; a shared entity map makes those links reusable.
Explore deeper if. you build agentic search or RAG over company documents.
What to doubt. Depends on accurate entity resolution across documents, which is hard in messy corpora.
Screened from the abstract, not tested. A reading guide, not a verification.
Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI
What the paper claims. A 'builder' model learns reusable principles ('meta-skills') for when another model needs support and what resources to give it, from that model's feedback, then builds harnesses for unseen tasks. Gains of about 9 points over building without skills and 12 over just handing the skills to the target model.
Why it matters if true. Shows that experience is more useful turned into executable support than into instructions, a step toward systems that improve their own environments.
Explore deeper if. you design agent environments automatically or study AI-for-AI loops.
What to doubt. Two benchmarks; development and test tasks may share structure.
Screened from the abstract, not tested. A reading guide, not a verification.
Traversing the solution space of neural networks with Hessian Null Space Continuation
What the paper claims. A method that walks through weight space along directions that keep the network's behavior unchanged, using local curvature. Near a single trained solution it finds networks with very different internal representations: in ViTs more different than independently trained models of other architectures, and in RL agents a different strategy that exposes reward hacking.
Why it matters if true. Interpretability findings about 'how a network computes' may describe just one of many equivalent mechanisms. It also opens tools for merging and editing.
Explore deeper if. you do mechanistic interpretability, model merging or study loss-landscape geometry.
What to doubt. Preserved behavior is measured on the training distribution; whether the variants behave the same off-distribution is a separate question.
Screened from the abstract, not tested. A reading guide, not a verification.
In-Context Learning for Robots: Methods and Applications
What the paper claims. A survey of in-context learning for robots: getting a fixed robot policy to do new tasks from demonstrations or interaction, without retraining. It groups methods into four families by how context reaches the controller and compares their assumptions and evaluation.
Why it matters if true. A map of a fast-growing area that clarifies which methods transfer under what conditions; useful before choosing an approach.
Explore deeper if. you are entering robot learning from demonstrations or want to compare ICL approaches for manipulation or navigation.
What to doubt. A literature review; it reports no new experiments.
Screened from the abstract, not tested. A reading guide, not a verification.
Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence
What the paper claims. Multimodal models that reason in hidden 'latent tokens' barely react when the image changes in ways that change the answer, because outcome-only rewards don't say which visual evidence matters. ReaLVR adds supervision tying latent tokens to relevant image evidence and improves results up to 235B models.
Why it matters if true. Latent reasoning is hard to inspect. This gives a test (do latent tokens respond to evidence?) and a training fix, relevant to anyone using hidden reasoning in vision models.
Explore deeper if. you train multimodal models with latent or continuous reasoning, or care about whether hidden reasoning is grounded.
What to doubt. Comparisons are with other latent-reasoning baselines; the abstract does not compare against ordinary text chain-of-thought.
Screened from the abstract, not tested. A reading guide, not a verification.
PanoVLN: Towards Effective Panoramic Vision-and-Language Navigation
What the paper claims. Simply giving a navigation model 360-degree images helps little. It helps once the model also plans longer action sequences, trains on routes with many branches, and combines semantic and geometric features. With RGB only it beats prior best success rates by about 9-12%.
Why it matters if true. Shows that more visual context needs matching changes in planning and training to pay off, a lesson that applies beyond navigation.
Explore deeper if. you work on vision-and-language navigation or robot navigation with wide-angle cameras.
What to doubt. Benchmark gains on R2R-CE and RxR-CE; real-world results are on one quadruped.
Screened from the abstract, not tested. A reading guide, not a verification.
What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling
What the paper claims. Robot world-action models that skip predicting the future at test time lose the generalization that justified predicting the future in the first place. Almost all of the benefit comes from the first denoising step, so a single forward pass on noised future video tokens (Simple-WAM) gives the generalization at near-latent speed.
Why it matters if true. Resolves a cost-versus-generalization debate in robot foundation models with a cheap recipe: prepare the future, don't fully generate it.
Explore deeper if. you build or deploy world-action models for robots and must trade inference cost against robustness.
What to doubt. Controlled comparisons use one backbone; how well the single-step finding transfers to other architectures is untested.
Screened from the abstract, not tested. A reading guide, not a verification.
Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence
What the paper claims. Represent a robot task's state and execution as code (objects, constraints, progress, recovery steps) that a coding agent writes and revises using feedback, calling learned policies as tools. Improves over a trained VLA on unseen composite tasks and completes physics experiments on a real dual-arm robot.
Why it matters if true. Makes robot behavior inspectable and fixable the way software is, and turns successful runs into reusable code and training data.
Explore deeper if. you combine coding agents with robot policies, or want explicit, revisable task structure for robots.
What to doubt. A broad vision paper; many 'self-evolution' claims are future work, and quantitative results cover a few benchmarks.
Screened from the abstract, not tested. A reading guide, not a verification.
Agent Priors-guided Policy Learning
What the paper claims. For robots learning skills from a few demonstrations, state each skill's structural assumption (for example, a grasp depends only on the gripper's pose relative to the object) both in training and in the skill's description. A planner can then pick and combine skills knowing where each works, which improves generalization and new combinations.
Why it matters if true. Connects how a skill is trained with how it is composed, a gap that causes skills to be used where they don't work.
Explore deeper if. you build compositional robot skill libraries from demonstrations.
What to doubt. Simulation benchmarks only (MetaWorld, ManiSkill) in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
LLMs are General Asynchronous Agents
What the paper claims. A framework that lets a standard LLM run several overlapping 'coroutines' with shared memory, so it can take in new inputs while thinking or acting. Qwen models handle streaming video, video games and monitoring this way without special training.
Why it matters if true. Voice assistants, robots and monitors need agents that react while working; this offers one general approach instead of a specialized architecture per use.
Explore deeper if. you build real-time or concurrent agents (voice, monitoring, games).
What to doubt. Demonstrations rather than benchmark comparisons, according to the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
From Input to Output: A Flexible Agent for Dual-End Interpretation of Sparse Autoencoder Features
What the paper claims. An agent that explains sparse-autoencoder features as a mapping from what activates them to what they do when changed, gathering evidence on demand instead of scanning a whole corpus. It improves both input and output explanation scores and finds that 71% of features mean different things on the input and output side.
Why it matters if true. Feature explanations that describe only inputs can mislead steering; this shows the two sides often differ and gives a cheaper way to explain both.
Explore deeper if. you label SAE features automatically or choose features for steering. Code is released.
What to doubt. Scores are from automated metrics; human validation of the explanations is not described in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Multilinguality in Hybrid Attention LLMs
What the paper claims. In hybrid models that mix full attention with recurrent layers, cross-lingual alignment spikes at the first full-attention layer. Every alternative layer ordering tested learns multilingual data faster than the standard one, up to 2.5x, suggesting multilingual models should start with full attention.
Why it matters if true. Layer ordering in hybrid models is usually copied without testing; this indicates a cheap architectural change that helps non-English languages.
Explore deeper if. you design hybrid-attention models or care about multilingual quality.
What to doubt. The ordering result comes from distillation experiments, not full pretraining.
Screened from the abstract, not tested. A reading guide, not a verification.
Less Sycophancy, Stronger Refusal? Lessons for AI Safety from Mechanistic Interpretability
What the paper claims. Training with a sycophancy feature injected (supplying the feature's activation during fine-tuning so the model doesn't learn it) cuts learned sycophancy by up to 62%. But less sycophancy does not reliably mean stronger refusal of harmful requests; the benefit appears only under user pressure, where it recovers most of the refusal lost to ordinary fine-tuning.
Why it matters if true. Shows an interpretability-based training intervention works, and warns against assuming that fixing one safety trait improves related ones.
Explore deeper if. you work on sycophancy, refusal robustness, or SAE-guided training interventions.
What to doubt. Three Qwen3.5 models; recovery numbers are for selected checkpoints, which risks optimistic selection.
Screened from the abstract, not tested. A reading guide, not a verification.
Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders
What the paper claims. Analyzing student models with a shared feature dictionary shows on-policy distillation neither creates new features nor copies the teacher's: it reweights features the student already has. The usual supervised warm-up does part of that reweighting in advance, and imposing the same reweighting without changing weights recovers most of the accuracy gain.
Why it matters if true. Changes the picture of distillation from 'transferring knowledge' to 'teaching the student to use what it has', which affects how to design and budget distillation.
Explore deeper if. you run on-policy distillation for reasoning or want interpretability tools for comparing training checkpoints.
What to doubt. Three settings; crosscoders may miss features they were not trained to represent.
Screened from the abstract, not tested. A reading guide, not a verification.
When Trees Are Not Enough: Learning Mixed-Topology Feature Graphs with Adaptive Graph Sparse Autoencoders
What the paper claims. A sparse autoencoder that learns relations between its features in which a feature can have zero, one or several parents, rather than forcing a tree, and uses that graph to guide training. It recovers mixed structures exactly in a toy model and gives more reliable relations and stronger causal control on real LLM activations.
Why it matters if true. Real concepts combine in non-tree ways; organizing SAE features accordingly could reduce feature absorption and make features more useful for control.
Explore deeper if. you study feature hierarchies or absorption in SAEs.
What to doubt. A complex training cycle; how it scales to very large dictionaries is not discussed in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
Raven: The Harness of Harnesses for Composable Agentic Intelligence
What the paper claims. Raven is an open-source multi-agent system that builds and improves its own specialized agent harnesses (the prompts, tools and workflows around a model) for each domain, then has a host agent split goals into subtasks and route them to the right model-harness pair.
Why it matters if true. Harness design is now a large part of how capable an agent is, and doing it by hand does not scale across domains. If harnesses can be built and composed automatically, cross-domain agents get cheaper to make.
Explore deeper if. you build multi-agent orchestration, or maintain several hand-tuned agent harnesses and want them generated or reused automatically.
What to doubt. 'Significantly outperforms' is not quantified in the abstract; the theory gives sufficient conditions for composition to help, not evidence that they hold in practice.
Screened from the abstract, not tested. A reading guide, not a verification.
Identifying Temporal Features within Transcoders for Time Sensitive Factual Recall
What the paper claims. Maps the individual features that let three small LLMs recall facts tied to a particular year, using transcoders. Time is represented through parallel, mixed syntactic and semantic features across layers, including higher-layer components that an existing circuit method misses.
Why it matters if true. Outdated or time-confused answers are a real failure; knowing the features involved enables targeted fixes instead of retraining.
Explore deeper if. you work on temporal knowledge, knowledge editing or transcoder-based circuits.
What to doubt. Models of 1-4B parameters and a specific recall task.
Screened from the abstract, not tested. A reading guide, not a verification.
Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL
What the paper claims. In RL for coding agents, all test-passing solutions get equal credit even if some make messy or unnecessary changes. A trained grader ranks passing solutions within each group and credit is shifted toward the better ones. Tested on 310B and 1T-parameter models, it improves performance, limits length growth and stabilizes training.
Why it matters if true. Rewards code quality, not just passing tests, at industrial scale, which is what users of coding agents care about.
Explore deeper if. you train coding agents with RL and see bloated or out-of-scope changes.
What to doubt. Depends on the grader's judgment of quality; the abstract gives no numbers for the gains.
Screened from the abstract, not tested. A reading guide, not a verification.
Omni-IO Skills: Harnessing Your Agent Omni-Native
What the paper claims. A plug-in harness that lets existing text-centric agents handle images, audio, video, documents, 3D and code through 27 'skills', a common execution interface and a registry of produced assets. It raises input support for two frontier agents from about 40% to 100% on the authors' benchmark.
Why it matters if true. Shows that broad multimodal capability can be added around an agent through tooling rather than retraining the model.
Explore deeper if. you build agents that must produce or edit many kinds of files and want an orchestration pattern for multi-asset workflows.
What to doubt. Evaluated on the authors' own benchmark (UniM-90); the gains largely come from making unsupported inputs possible at all, not from better reasoning.
Screened from the abstract, not tested. A reading guide, not a verification.
Understanding the Role of Prompt Template in Knowledge Distillation for Safety Alignment
What the paper claims. When distilling a teacher into an already safety-aligned student, using the chat template during distillation makes the student noticeably more willing to answer harmful queries than using a plain template. The chat template also shifts internal representations more. Consistent across Llama, Gemma and Qwen.
Why it matters if true. A simple, overlooked setting can quietly erode safety during distillation; easy to fix once known.
Explore deeper if. you distill into instruction-tuned models that must stay safe. Code is released.
What to doubt. Effect sizes are not in the abstract; it is unclear whether task quality differs between template choices.
Screened from the abstract, not tested. A reading guide, not a verification.
Post-Training Leaves Behavioral Shadows on Unrelated Decisions
What the paper claims. A student model can pick up a teacher's post-training capability (for example +5.3 points on a coding benchmark) from nothing but the teacher's choice of one ordinary word per prompt, on prompts where their shared ancestor was undecided between two words. No task examples, logits or weights are used.
Why it matters if true. If capability leaks through such tiny, task-unrelated signals, model outputs carry far more transferable information than assumed. That matters for distillation, for data contamination and for protecting models from being copied.
Explore deeper if. you work on distillation, subliminal learning, model IP protection, or the safety of training on other models' outputs.
What to doubt. Requires teacher and student to share a common ancestor; the main result is one coding experiment on a 1.5B model (others are summarized), and the gain is modest.
Screened from the abstract, not tested. A reading guide, not a verification.
Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
What the paper claims. If you feed a transformer a linear mix of two different text streams, its output is close to the average of the two separate next-token predictions. This seems built into the architecture, weakens during pretraining, and can be restored with light fine-tuning, which allows generating two continuations in one forward pass.
Why it matters if true. A surprising linearity in a nonlinear model, with a possible use: two generations for the price of one.
Explore deeper if. you study transformer linearity or superposition, or look for unusual ways to batch generation. Toy-checkable on small models.
What to doubt. The abstract does not report how close the outputs are, at which model sizes, or the quality of the disentangled generations.
Screened from the abstract, not tested. A reading guide, not a verification.
Small yet Assistive: Spatially-Aware Post-Training for Low Vision
What the paper claims. A 500M vision-language model for blind and low-vision users, post-trained with distillation and an RL reward for directions, distances and hazards, plus a final step to recover general description quality. It runs offline on a mid-range Android phone and improves spatial and OCR scores substantially.
Why it matters if true. Brings navigation-grade scene description to phones without internet, for a population of about a billion people. Model, data and code are released.
Explore deeper if. you build accessibility tools or need small, on-device vision-language models.
What to doubt. Gains are relative to the small baseline; real-world testing with blind users is not described in the abstract.
Screened from the abstract, not tested. A reading guide, not a verification.
The System Prompt Illusion: How Instruction Preambles Modify Computation in Language Models
What the paper claims. Across 17 models from 1.5B to 72B, persona and formatting system prompts reshape internal representations, while safety instructions barely move them. Restrictive and 'you have no restrictions' prompts produce nearly identical computation. Models register the prompt but don't act on it deeply.
Why it matters if true. A mechanistic explanation for why system-prompt safety is easy to jailbreak: relying on system prompts for safety is weaker than it looks.
Explore deeper if. you rely on system prompts for safety or behavior control, or study instruction following. Code is released.
What to doubt. Representation similarity is an indirect measure; the link to behavior is supported by patching and a cross-model correlation.
Screened from the abstract, not tested. A reading guide, not a verification.
SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
What the paper claims. Memory for group chats that keeps two tracks: speaker-labeled verbatim messages, plus derived person-level and group-level states, combined by entity, event and time at query time. A small writer model trained with RL improves accuracy, giving the best reported score on one leaderboard.
Why it matters if true. Group assistants must track who said what about whom; general memory systems confuse speakers and relationships.
Explore deeper if. you build assistants for multi-person chats (team tools, community bots).
What to doubt. Accuracies of 48-69% show the task remains hard; gains are on recent, specialized benchmarks.
Screened from the abstract, not tested. A reading guide, not a verification.
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
What the paper claims. Automatically evolving an agent's harness tends to overfit the training tasks. Adding regularization (limit how many edits each candidate may bundle, screen out benchmark-specific changes, prune useless ones) keeps up to 4.3 points of gain on unseen benchmarks and uses 30% fewer tokens.
Why it matters if true. Automated harness tuning is spreading; this brings the basic machine-learning lesson of regularization to it, so gains survive outside the tuning set.
Explore deeper if. you auto-tune prompts, workflows or harnesses and worry that gains don't hold up on new tasks. Code is released.
What to doubt. Gains are a few points; whether the critic can reliably tell benchmark-specific edits from general ones is a judgment the abstract does not detail.
Screened from the abstract, not tested. A reading guide, not a verification.
Muon Can Outperform Dedicated Continual Learning Methods
What the paper claims. Plain incremental LoRA trained with the Muon optimizer matches dedicated continual-learning methods on one benchmark and beats every AdamW setup on another. The difference is not update size but spread: AdamW concentrates updates in 1.4-1.8 effective directions, Muon spreads them over about 7.
Why it matters if true. Part of what continual-learning methods are credited for may simply come from optimizer geometry. Switching optimizer could be a free baseline that methods should beat.
Explore deeper if. you do continual fine-tuning with LoRA or evaluate continual-learning methods. Small-scale reproducible.
What to doubt. Two benchmarks with 3-5 seeds; whether Muon's advantage holds at larger scale and against newer methods is open.
Screened from the abstract, not tested. A reading guide, not a verification.
Grounded Action Model: 3D Grounding as a Foundation for Robotics
What the paper claims. Robot foundation models built around 3D object grounding: prompts (language, points, boxes) become object-centric features with metric geometry, mixed with robot state to predict actions. More robust to scene changes than leading policies, for example 17/20 vs 4/20 successes under visual shift on a real robot.
Why it matters if true. Supports the idea that explicit 3D grounding, not language or video pretraining alone, is what robot policies lack.
Explore deeper if. you design robot policy architectures or need robustness to scene and target changes.
What to doubt. Some margins on simulated benchmarks are small (55.3% vs 52.0%).
Screened from the abstract, not tested. A reading guide, not a verification.
Machine-Interpretable Information: Compiling Documents into Searchable and Readable Protocol States
What the paper claims. Compiles each document into a fixed 56-token state that any of several different LLMs can read through a small translator, serving as a search key, a memory and a source for reconstruction. With added sparse local evidence it beats full-context answers on HotpotQA using about 7% of the attention compute.
Why it matters if true. A transferable 'compiled' document format could cut long-context and RAG costs across models.
Explore deeper if. you run RAG at scale and pay for re-reading long documents, or work on context compression.
What to doubt. Results on one QA dataset; fixed 56-token bandwidth may lose detail for long or dense documents.
Screened from the abstract, not tested. A reading guide, not a verification.
OmniEdu: Open Foundation Models for Learning and Teaching
What the paper claims. An open family of 4B, 9B and 27B models fine-tuned on about 70,000 curated examples organized by teaching capability (subject knowledge, curriculum, diagnosing mistakes, scaffolding). They improve on K-12 problem solving, curriculum and tutoring benchmarks.
Why it matters if true. Open education-specific models with a documented data pipeline, useful for tutoring products and for research on what data teaching ability needs.
Explore deeper if. you build K-12 tutoring or education tools and want an open base model or a data recipe.
What to doubt. Gains are on benchmarks; the abstract does not report classroom or learner outcomes, or how much general capability changed.
Screened from the abstract, not tested. A reading guide, not a verification.
Transferring the Intelligence of VLMs to Robotic Control
What the paper claims. A general vision-language model controls a robot through a small set of discrete move, rotate and gripper commands in a closed loop. Zero-shot it beats several policies trained on the benchmark's own robot data, and one demonstration in context raises success from 53% to 74%, above a strong trained baseline.
Why it matters if true. Suggests that digital-world intelligence transfers to robot control with the right interface, lowering the need for robot-specific training data.
Explore deeper if. you want to use frontier VLMs for manipulation without collecting robot data, or compare interfaces for VLM control.
What to doubt. Discrete step commands are slow; real-robot tests are simple block tasks.
Screened from the abstract, not tested. A reading guide, not a verification.
Also new: who should care (183)
artist: A Python package for differentiable ray tracing and gradient-based optimization of solar tower power plant operation
Who should care: Solar-thermal engineers wanting a differentiable ray-tracing simulator to optimise heliostat fields, not an ML-method claim.
Deep Learning for Physicists: From Linear Regression to Reasoning Models
Who should care: Physicists or newcomers wanting lecture notes that take deep learning from linear regression to reasoning models.
A COMPASS to the Galaxy: Joint nucleosynthetic yield set comparison and parameter inference in Galactic chemical evolution
Who should care: Astrophysicists comparing stellar nucleosynthesis yield sets while inferring Galactic chemical evolution parameters.
Abstention Direction in Large Language Models: Mechanistic Interpretability of Abstention Across Query Ambiguity, Knowledge Limits, and Safety Alignment
Who should care: Alignment researchers asking where in the model an abstain-or-answer decision sits across ambiguous, knowledge-limited and safety cases.
Certified, Not Explained: Exact Token Erasure in Gated Vision State Space Models
Who should care: Medical-imaging explainability users who rely on a Mamba model's step size to show which patches mattered; claims it can mislead.
GPRM: Global Perspective Process Reward Model
Who should care: People building step-level reward models for multi-step reasoning who need better history and global context in scoring.
AsymEEF: Bimanual Manipulation with Asymmetric End Effectors
Who should care: Bimanual robotics groups wanting to reuse demonstrations across end-effectors with different kinematics.
SWE-OPD: On-Policy Distillation for Agentic Coding
Who should care: Teams distilling agentic coding ability from a big teacher into a smaller student on-policy.
Generative AI-Driven Parametric Design Optimization Framework for Vehicle Closure Systems
Who should care: Vehicle design engineers wanting generative design tied to editable CAD and downstream analysis.
DexFormer 2.0: Towards a Cross-Embodiment Foundation Policy for Dexterous Grasping
Who should care: Robotics teams who want one dexterous-grasp policy across many arm and hand pairings instead of one policy per pairing.
GEM-MM: Reducing VLM Hallucination with Entropy-Guided Preference Alignment
Who should care: Vision-language alignment teams with scarce preference data who need token-level entropy guidance against hallucination.
Ego2Sim2Humanoid: Agentic Real-to-Sim for Learning Loco-Manipulation VLAs without Robot Demonstrations
Who should care: Humanoid robotics groups wanting to train loco-manipulation policies from egocentric human video through simulation instead of teleoperation.
MARS: Dynamic Routing of Reinforcement-Learned Atomic Skills for Multi-Stage Robotic Manipulation
Who should care: Robot manipulation teams who need precise grasp and alignment stages by routing among learned atomic skills.
Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination
Who should care: Agent builders debugging web-element hallucination who want to know if interpretability can explain agent failures.
Ego2Flow: Benchmarking the Traffic-System Impact of Autonomous Driving on Maps Distilled from Real-World Roads
Who should care: Autonomous driving evaluators interested in the traffic-system effects of end-to-end driving, not only ego safety.
MemeEconomy: Do LLM Agents Trade Ethics for Survival?
Who should care: Safety researchers testing whether aligned LLM agents trade ethics for survival in multi-round settings.
NER-D: A cross-domain benchmark for German NER
Who should care: German NLP teams needing a cross-domain named-entity benchmark across nine domains.
A Multi-Layered AI Pipeline for BPMN Generation from Natural Language
Who should care: Business-process teams wanting natural-language descriptions turned into BPMN diagrams.
LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent
Who should care: Data teams wanting a fast one-pass text-to-SQL model in front of a slower agent, for privacy and cost.
NeoResearch: Literature- and Experience-Guided Autonomous Time Series Model Discovery
Who should care: Forecasting teams wanting an autonomous agent that discovers time-series models from literature and experience.
Explanations that Transfer: Interpreting Medical Retrieval with Sparse AutoEncoders
Who should care: Medical retrieval teams wanting sparse-autoencoder explanations that transfer across fine-tuned embedders.
Uncertainty-Aware Continual Learning for 2D XRD Interpretation Using Simulated and Experimental Data
Who should care: Materials scientists running autonomous X-ray diffraction labs who need models that stay calibrated as new experimental data arrives.
Parameter-Efficient Conciseness Adaptation: Differentiable Length Control for Concise Clinical Reasoning in Smaller Language Models
Who should care: Clinical NLP teams deploying small models who need concise outputs through a differentiable length-control adapter.
EvenVLM: A Text and Multimodal Dataset for the Even Language with Vision-Language Fine-Tuning
Who should care: Researchers on endangered languages wanting a text and visual question dataset for Even.
Lara Mini Base: Language Diversity as a Resource, Not a Trade-off
Who should care: Teams building multilingual base models who argue language diversity helps rather than hurts.
Alignment with Reflect-Reasoning: Learning How To Follow Principles
Who should care: Alignment teams who need models to follow context-dependent principles by reflective reasoning.
Interpretable machine learning diagnosis of degradation mechanisms in lead-free $RbSn_{0.5}Ge_{0.5}I_{3}$ solar cells from temperature- and illumination-dependent electrical measurements
Who should care: Solar-cell researchers diagnosing degradation mechanisms from device measurements with interpretable ML.
Climate Temporal Temperature Prediction via an Adaptive Stacked Bidirectional Residual Mamba Network
Who should care: Weather and climate forecasters considering a stacked Mamba network for temperature prediction.
A Mechanistic Interpretability Study of an Astronomy Foundation Model
Who should care: People asking whether a scientific foundation model computes known astronomy quantities, using interpretability on an astronomy model.
Where Is Uncertainty Observed? A Mechanistic Understanding of Uncertainty Quantification in LLMs
Who should care: Teams building LLM uncertainty estimators who want to know at which internal point uncertainty shows up.
ReShap: Interpretable Reinforcement Learning for High-Dimensional Multi-Action Settings
Who should care: Reinforcement-learning practitioners wanting interpretable policies for high-dimensional multi-action control.
The Illusion of Diversity: Aligning LLM Exploration via Effective Entropy
Who should care: People training reasoners with verifiable rewards who see diversity collapse and want an effective-entropy view of exploration.
Actuation and Control of Passive Tendon-Driven Soft Extensions Through Rigid Manipulator Motion
Who should care: Soft-robotics engineers wanting to actuate passive tendon-driven soft extensions by moving a rigid arm.
Taxonomy-Agnostic Neurosymbolic Grounding of Robot Kinematics
Who should care: Robotics teams linking task planners to hardware: converts URDF robot descriptions into ontologies using an LLM and graph purging.
From Egocentric Human Video to Robot Manipulation in the Era of Embodied Intelligence
Who should care: Robot-learning researchers: a survey of using egocentric human video for manipulation when no robot actions are recorded.
Mechanistic Interpretability of Deep Neural Networks: Methods, Frameworks, and Perspectives
Who should care: Newcomers to interpretability: a survey of methods and frameworks across image, audio and language networks.
Sophon: A Procedural Diagnostic for Spatial Reasoning in Vision-Language Models
Who should care: Vision-language model evaluators who want a procedural test of spatial reasoning with object and texture priors removed.
Conformal Risk Certificates for Portfolio Optimization Under Uncertainty: A Gated Reinforcement-Learning Approach
Who should care: Finance and risk teams wanting a replayable risk-bound certificate before executing a reinforcement-learning portfolio allocation.
Large Language Models as a Universal Searcher for Open-World Automated Machine Learning
Who should care: AutoML users: an LLM trained on search trajectories from one vision task that tunes architecture, hyperparameters and data ratios across modalities.
CADMA: Capacity-Aware Coverage Decomposition for Generative Model Assessment
Who should care: Generative-model evaluators: a recall-style metric that splits coverage of the real distribution while accounting for model capacity.
Detecting and Explaining Robot Policy Failures from Sparse Internal Features
Who should care: Robot-policy developers: reads a policy's sparse-autoencoder features to flag failures during execution and name the internal concepts involved.
AERO: Adaptive Ensemble-Disagreement Routing for Oracle Feedback in Sample-Efficient Online RLHF
Who should care: RLHF practitioners with a labeling budget: routes preference queries to an oracle based on disagreement among an ensemble of reward models.
Learning a Unified Cross-Model Semantic Dictionary via Gated Bottleneck Sparse Autoencoders
Who should care: Cross-model interpretability researchers: a shared sparse-autoencoder dictionary meant to match features across different foundation models.
SWEET: Mitigating Spurious Feature Reliance in LLM Reasoning via Weight Editing
Who should care: Teams troubleshooting brittle LLM reasoning: edits weights to suppress spurious features that make chain-of-thought fail on distractors.
Cross-Layer TopK: Dynamically Allocating Sparsity Across Layers in Transcoders
Who should care: Circuit-tracing researchers: cross-layer transcoders that allocate sparsity per layer to train more efficiently.
KnowEvo: Knowledge Evolution for Protein Optimization
Who should care: Protein-engineering teams using LLMs: accumulates and reuses scientific knowledge across optimization rounds.
User Embeddings are Superpositions of Interpretable Behavioral Modes
Who should care: Recommender-system teams: reads dense user embeddings as sparse mixtures of interpretable behavior modes.
Curvature by Design: A Support Structure for Continuum Robots
Who should care: Continuum-robot designers: a support structure that fixes the curvature profile, instead of assuming it in the kinematics model.
Shaping & Monitoring Models through their Internal Representations
Who should care: Alignment teams: trains models so internal representations follow rules about which concepts they express, and monitors them.
The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention
Who should care: Teams fine-tuning with DPO on narrow tasks: an empirical look at which capabilities drop (Qwen3.5-4B, financial filings).
PEARL: Prediction Error-Adaptive Regularized Learning for Selective Forgetting under Partial Preference Drift
Who should care: Recommender and RLHF teams facing preference drift: forgets only the dimensions that shifted.
TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport
Who should care: LLM safety researchers: blocks jailbreaks by tracing and realigning cross-layer safety circuits with sparse autoencoders, not by ablating single neurons.
Modulating Merging Strengths via Joint Loss Estimation for LoRA-based Continual Learning
Who should care: Continual-learning practitioners using LoRA: merges each new adapter with per-element strengths solved from a joint loss on old and new tasks.
SaMA: Morpho Adaptation via Asymmetric Expansion of Kronecker Product
Who should care: PEFT researchers: SaMA replaces LoRA's factorization with an asymmetric Kronecker form to remove flat directions in the loss.
CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
Who should care: Quantitative-finance researchers: LLM-driven search for alpha factors over a multi-relational graph.
Replay for LLM Continual Learning with Multi-Teacher On-Policy Distillation
Who should care: Post-training teams merging domain experts over time: replay for multi-teacher on-policy distillation when experts arrive one after another.
Alignment Flywheel-Guided Suppression Patching for Interactionless Inverse Reinforcement Learning
Who should care: Inverse-RL researchers: patches a learned reward model so policies cannot exploit its errors.
Decomposing Irregular Time Series into Interpretable Temporal Patterns with Mixture-of-Experts
Who should care: Finance and logistics analysts with irregular logs: an interpretable mixture-of-experts that splits transaction series into temporal patterns.
A Framework for Task-Agnostic Counterfactual Fairness in Transformers through Activation Steering
Who should care: Fairness auditors of transformers: uses activation steering to reduce counterfactual bias without task-specific training.
Mechanistic Interpretability for End-to-End Self-Driving
Who should care: Autonomous-driving safety engineers: sparse autoencoders on a single-camera path planner to see what features drive its plans.
SketchMate: Agentic Geometric Reasoning for Zero-Shot Precision Insertion
Who should care: Robot assembly teams: a vision-language model builds a geometric sketch to insert unseen keyed parts with sub-millimetre accuracy.
From Concepts to Actual Causes: Toward Closing the Concept-Malfunction Gap in Robot Foundation Models
Who should care: Robot incident investigators: asks which internal components of a foundation-model policy caused a failure.
MUSBRIC: LEGO-Compatible Pneumatic Muscle Bricks for Reconfigurable Musculoskeletal Robots
Who should care: Robot-design hobbyists and researchers: LEGO-compatible pneumatic muscle bricks plus a MuJoCo simulator for quickly rebuilding musculoskeletal robots.
Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI
Who should care: Teams choosing open models for agents: NVIDIA's 120B and 30B mixture-of-experts Mamba-Transformer models with Indic pretraining and agentic coding training.
Watermark Basins: Black-Box Detection of Distilled LLMs
Who should care: Model providers worried about distillation: black-box watermark detection of whether a student was trained on a teacher's outputs.
DocTalk: Enhancing Medical Knowledge Retrieval via Fine-tuned Large Language Models with Chain-of-Thought Reasoning for Robust Clinical Decision Support
Who should care: Hospital and clinical-AI groups: a LoRA-tuned question-answering system for discharge summaries, tested on 500 generated QA pairs.
Structure, Not Sequence: Tracking and Validating Folding Mechanisms in ESMFold's Trunk with Sparse Autoencoders
Who should care: Protein-ML researchers: sparse autoencoders on ESMFold's trunk to see how structure is computed from sequence.
HIDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning
Who should care: Lifelong imitation-learning researchers: hierarchical routing attention that avoids storing past data.
HumanoidNDM: Closing the Training-Deployment Gap of Motion Tracking via Humanoid Neural Dynamics Model
Who should care: Humanoid motion-tracking teams: a neural dynamics model to narrow the gap between simulation and real-world deployment.
Large Language Models: A Plain-Language Overview for Engineers
Who should care: Engineers new to LLMs: a plain-language overview of what training and prompting each change.
Intrinsic Goals for Autonomous Agents: Hierarchical Exploration Captures Future Volitional Action in Drosophila
Who should care: Neuroscience and agent researchers: links fruit-fly whole-brain recordings to hierarchical intrinsic goals in artificial agents.
Speaking-Clock Mamba: Phoneme-Conditioned State-Space Modeling for Visual Forced Alignment
Who should care: Speech and video alignment teams locating phoneme boundaries in silent talking-face video.
MOTIO: Agentic Skill Engineering for Reality-Grounded Humanoid Loco-Manipulation
Who should care: Humanoid-robotics groups: an agent system that turns human videos and 3D assets into staged loco-manipulation programs.
Concept Acupuncturist: An Efficient and Precise Concept Erasure Method for Diffusion Models via Sparse Autoencoders
Who should care: Safety teams removing concepts from image diffusion models: uses sparse autoencoders to find the one cross-attention block that holds a concept and edits only it.
Propensities Please: Exposing Alignment-Faking LLMs using Synthetic Document Finetuning
Who should care: Alignment evaluators: tries to expose models that hide misaligned goals by fine-tuning on synthetic documents instead of relying on prompts or chain-of-thought.
Sharpness-aware Orthogonalized Momentum Algorithms
Who should care: Optimizer researchers: adds sharpness-aware minimization to Muon-style orthogonalized momentum.
Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook
Who should care: Anyone starting on time-series anomaly detection: a survey with a taxonomy, metric-reliability experiments and a benchmark repository.
TaxoSAE: Rethinking Sparse Autoencoders through Taxonomic Logic
Who should care: SAE researchers: makes sparse-autoencoder features follow a taxonomy of concepts instead of being independent units.
HERD: Harness-Evolving Recursive Distillation for Continual Learning
Who should care: Agent builders who need continual learning: distils an evolving harness of prompts and skills back into the model.
SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs
Who should care: Agent-safety engineers: compiles natural-language policies into checkable programs over multi-turn transcripts.
Formally Verified Autoresearch for Analyzing the Expressivity of Multi-Head Attention
Who should care: Theory readers interested in how many attention heads a one-layer transformer needs, with machine-checked proofs.
BatchParScale: Sparse Batch-Mixing for Sequence Models
Who should care: Inference teams doing parallel sampling who want streams to mix information instead of only voting at the end.
A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help
Who should care: Multimodal model builders asking whether vision-language training harms how the language model represents text.
Quality-Controlled Reply and Summary Data for Post-Training Small Language Models in AI Call-Screening Agents
Who should care: On-device assistant builders needing quality-controlled data to post-train small models for call screening.
PACT: NVFP4 Data-Flow LoRA Fine-Tuning on Consumer GPUs
Who should care: People fine-tuning LoRA on consumer GPUs who want lower memory through NVFP4 data flow.
SurGPT: A Multi-Task Foundation Model for Robotic Surgical Skills Assessment and Gesture Recognition
Who should care: Surgical education teams wanting automated robotic-surgery skill scoring and gesture recognition.
TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis
Who should care: Builders of physics-reasoning training data who need verifiable generated problems.
HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
Who should care: Deep-research agent builders whose questions span both tables and free text.
First-Order LoRA Training with Phone NPUs
Who should care: Mobile developers wanting on-device LoRA fine-tuning on phone NPUs for private data.
Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning
Who should care: Retrieval-augmented system builders wanting learned compositions of search primitives instead of fixed pipelines.
VISTA: Steering LLMs with Latent User Values Inferred from a Few Survey Answers
Who should care: Personalization researchers: infers a user's values from a few survey answers and steers the model's activations with them.
GRASP: Global-Local Representation Alignment and Spatial-Physical Grounding for Bimanual Robotic Manipulation
Who should care: Bimanual robot-manipulation teams: training losses that weight joint errors by their physical consequences and focus on hand-object regions.
Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
Who should care: Chain-of-thought faithfulness researchers: describes in words how reasoning anchors move through layers, to detect unfaithful reasoning.
Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism
Who should care: Sycophancy and bias researchers: traces several prompt-cue biases to one attention circuit that already exists before post-training.
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning
Who should care: Efficient-reasoning researchers: uses separator tokens as anchors to compress chain-of-thought into latent steps.
OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
Who should care: Image-generation and world-model teams: a verifier that grades each part of a compositional prompt or rule instead of a single pass/fail.
First Passage Attention
Who should care: Sequence-architecture researchers: an attention variant that can erase intermediate context while keeping older relevant tokens.
Noise-shaped Sampling and Annealing for Scalable On-the-edge Training of Binary and Ternary AI Models
Who should care: Edge-AI hardware researchers: trains binary and ternary networks with Ising-style solvers suited to on-device training.
Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
Who should care: Generative-model teams wanting diverse outputs: mines a category scheme from real data and samples from it at test time.
GRID-Planner: Graph-Structured Relational Reward Shaping for Diffusion Planning
Who should care: Autonomous-driving planning teams: reward shaping from the relations between road agents, for diffusion-based planners.
Risk and Refusal: Mechanisms of LLM Responses to Suicidal Ideation
Who should care: AI-safety and mental-health teams: how LLMs respond to suicidal-ideation disclosures and what internal mechanisms drive refusals.
Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures
Who should care: Agent reliability teams: treats human interventions as labels to locate, steer and sometimes edit the computation behind an agent failure.
Emotion Is Not External Knowledge: Internal Attention Signals for Affective State Reasoning
Who should care: Dialogue researchers: finds attention signals that track a user's emotional state across topic shifts in multi-turn chats.
Designing Textual Interfaces To Enhance Llm Reasoning On Audio-Visual Question Answering
Who should care: Multimodal QA teams using prompt optimization: how to expose audio-visual evidence as text so optimizers can improve the prompts.
Safe-FCL: Continual Risk Awareness and Fine-Grained Safety Aggregation for Federated Continual Learning of LLMs
Who should care: Federated-learning security researchers: detects malicious client updates to LLMs as attack patterns change over time.
FACES: Fine-grained Gated Semantic Alignment for Multimodal Emotional Support Conversation
Who should care: Emotional-support chatbot researchers: multimodal response generation with finer-grained reward signals.
PRISMATIC : Prescription Risk Inspection System for Multi-Agent Treatment Interaction in Clinical Decision-Making
Who should care: Clinical decision-support teams: multi-agent checking of prescriptions for patients with multiple conditions.
AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive Coding
Who should care: Engineers running long-lived agents: compresses persistent KV caches under a set reconstruction-error budget.
Explore with Memory, Learn without It: In-Context Exploration and Self-Distillation for LLM Reasoning
Who should care: RL-for-reasoning teams: puts memory of earlier failed attempts into the prompt during exploration, then distills so the final model needs no memory.
TSRL: Trigger-based Supervised Reinforcement Learning for Continual Agent Training
Who should care: Teams continually training LLM agents: combines supervised fine-tuning and RL to learn new skills with less forgetting.
Monitoring latent reasoning in looped language models
Who should care: Researchers on looped (weight-reused) language models: checks whether extra loops actually improve answers and how stable they are.
Narrowing the Collaboration Gap, Probably
Who should care: Multi-agent LLM system designers: agents share probability estimates rather than only proposed actions, to pool information better.
Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents
Who should care: Long-horizon agent builders: keeps a belief state across attempts to trace failures back to earlier causes.
MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation
Who should care: Vision-language navigation researchers: compresses and reuses spatial memory across long navigation trajectories.
Walk Like This: Stylized Locomotion for Physics-Based Characters through Example-Driven Diffusion
Who should care: Character-animation and game teams: real-time physics-based locomotion with style taken from examples, without style labels.
inLoRA-MINT: Internal Representations for Auditing Training Data in Domain-adapted LLMs
Who should care: Privacy and data-audit teams: tests whether specific samples were used to fine-tune a LoRA model, using its internal activations.
Act4R: Action-conditioned 4D reconstruction for kinematics-free control
Who should care: Robot-control researchers: learns how a robot's body moves from video and actions, without a kinematic model.
UZHCL_for_the_win at Touche: Encoder Fine-Tuning, Synthetic Augmentation, and an Underpowered Encoder-LLM Comparison
Who should care: Argument-mining and fallacy-detection researchers: a shared-task system on Reddit data comparing encoder fine-tuning with LLMs.
Neural Bayesian Networks for Modular and Scalable Decision-Making Under Uncertainty
Who should care: Decision-making-under-uncertainty researchers: Bayesian networks with neural conditional distributions for larger, modular problems.
MLAfication: Efficiently Enabling High-Fidelity MLA for Modern LLMs
Who should care: LLM inference engineers: converts existing models with QK-norm, including MoE, to multi-head latent attention to shrink the KV cache.
Welcome to the Age of Agent Ads: Using On-Policy Attacks to Steer Agentic Decisions
Who should care: AI-agent security researchers: shows how on-policy 'agent ads' can steer agents' rankings for shopping, jobs and travel.
Event Memory and Narrative Dynamics: A Large Language Model Framework for Multidimensional State Spaces
Who should care: Memory and narrative psychologists: an LLM framework for segmenting autobiographical narratives into events and modeling their structure.
Differential Privacy Protected Preference Alignment on Survey Derived Data
Who should care: Privacy-preserving alignment researchers: preference alignment with differential privacy on sensitive survey data.
LoongSage: A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes
Who should care: Teams building agentic RL training infrastructure: a production RL framework with validated recipes for frontier-scale models.
An Explainable System for Offensive Content Detection in Nastaliq Urdu: Corpus Curation, Transformer Benchmarking, and Word-Level Attribution
Who should care: Content-moderation teams working in Urdu: a new Nastaliq-script corpus, transformer benchmarks and explanations for offensive-content detection.
LLM Value Functions for Credit Assignment in Agentic RL
Who should care: Agentic-RL teams: uses an LLM as the value function to give individual actions credit in long, sparse-reward tasks.
CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation
Who should care: Medical-imaging AI teams generating radiology reports: disease-aware image-text alignment producing structured, QA-style reports.
REFLECT: Role-Aware Detection of Sentiment, Emotion, Politeness, and Formality
Who should care: Dialogue and affective-computing researchers: detects sentiment, emotion, politeness and formality using explicit semantic roles.
DyssomniaNN: Multi-Scale Spiking Neural Modeling of Human Sleep Disorders and Perceived Restedness
Who should care: Computational sleep neuroscientists: a spiking-network model of why sleep disorders feel different from what sleep measurements show.
Proactive Agentic Orchestration
Who should care: Builders of multi-agent LLM systems: orchestrators that can revise sub-agents' goals mid-task and learn from denser feedback.
Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents
Who should care: LLM agent builders working on long tasks: lets agents rewind to earlier states and carry lessons from abandoned attempts forward.
X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation
Who should care: Dataset-distillation researchers in audio-visual learning: smaller synthetic datasets that keep cross-modal correspondence with less memory.
AgentTune: A Unified Training Lifecycle for Tool-Using LLM Agents
Who should care: Teams training tool-using LLM agents: a source-available library covering building, training, evaluation and distillation in one place.
Sparse Dictionary Learning for Measure-Valued Data via Linearized Optimal Transport: Efficient Algorithms and Sample Complexity
Who should care: Optimal-transport and representation-learning researchers: sparse dictionary codes for data that are probability distributions.
OpenPA: Forced-alignment-free Multi-granular Pronunciation Assessment with Mixed-transcript Adaptation
Who should care: Speech and language-learning technology teams: pronunciation scoring from Whisper without forced alignment.
Tracing and Taming Hallucinated Terms via Subsequence-Association Filtering and Breadth-Aware Selection for LLM Query Expansion
Who should care: Search and retrieval engineers using LLM query expansion: traces and filters hallucinated expansion terms.
Retractions Revisited: A Replication, Extension, and Mechanistic Evaluation of Linguistic Obfuscation Effects in Fraudulent Science
Who should care: Research-integrity and meta-science researchers: replicates the claim that fraudulent papers use more obfuscating language, on 1,600 papers.
From Structure to Validity: Mechanism-Guided Generation and Refinement of LLM Rubrics
Who should care: Teams evaluating LLMs with rubrics: which rubric-design choices (structure, specification, aggregation) actually improve evaluation validity.
SharP: Soft and Hard Prompt-Guided Augmentation with LLM for Low Resource Fake News Detection
Who should care: Misinformation-detection researchers with little labelled data: LLM prompt-guided data augmentation for fake-news classifiers.
JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models
Who should care: AI-for-science and autonomous-research builders: forecasts experiment outcomes with a world model to cut costly trial-and-error.
Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
Who should care: Robot-learning researchers fighting forgetting: continual steering of frozen generalist policies using world-model rehearsal.
VQIL: Visible Quantum Interference Learning for Perturbation Response Ranking on Graphs
Who should care: Network-science and quantum-ML researchers: ranks which graph nodes respond most to a perturbation using a trainable quantum walk.
Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning
Who should care: Chemistry-AI teams: tool-using chemical reasoning agents built on masked diffusion language models instead of autoregressive ones.
Monarch-MoE: Structured Experts for Multimodal Continual Instruction Tuning
Who should care: Multimodal continual-learning researchers: block-structured LoRA experts that isolate updates to limit forgetting.
Personalized Federated Direct Preference Optimization via LoRA Adapters
Who should care: Federated and personalized alignment researchers: per-client LoRA adapters for DPO across clients with different preferences.
TreeHCA: Hindsight-Guided Exploration and Credit Assignment for Multi-Turn LLM Agents
Who should care: RL-for-LLM-agents researchers: finer credit assignment than GRPO for multi-turn agents via hindsight-guided exploration.
MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures
Who should care: Embodied multi-robot agent researchers: a benchmark for long tasks with changing goals and execution failures.
PartiGen: Part-Aware Representations for Language-Conditioned Humanoid Motion Generation
Who should care: Humanoid-robotics and motion-generation teams: text-to-motion with body-part-aware diffusion suitable for robot execution.
PATACON: GPU-Parallel Tangent-Bundle RRT-Connect for Constrained Motion Planning
Who should care: Robot motion-planning engineers: GPU-parallel constrained planning for humanoids fast enough for continuous replanning.
Sampling-based Kinodynamic Planning with World Models
Who should care: Robot-learning researchers planning with visual world models: tree search in latent space for long-horizon manipulation.
Examining Social Attribution in LLM Reasoning: A Theory-Guided Probing Methodology
Who should care: Researchers auditing how LLMs assign blame or responsibility in social scenarios; uses probing grounded in social-psychology theory.
QUILT: Rethinking Sparse-Attention Prefill through Shared Query Execution
Who should care: LLM inference engineers speeding up long-context prefill: shares KV loads across neighboring queries in sparse attention.
UXBench Pro: Benchmarking Personalized User Experience in Multi-Turn Dialogue Interactions
Who should care: Teams evaluating chat assistants per user rather than with one reward model: a 1,000-case personalized UX benchmark.
Beyond Sequences: Distilling Structured Decision Memory for LLM Recommendation
Who should care: Recommender-system teams using LLMs on multi-behavior data (views, carts, purchases) who want more structure than flat token sequences.
Fact over Fiction: Detection of Pathological Hallucinations in Sinhala-to-English Neural Machine Translation
Who should care: Machine-translation practitioners in low-resource languages who need to detect hallucinated translations without reference translations.
From a Prompt to Repertoires: Evolving Functional REpertoires Enable LLM Continual Learning
Who should care: Continual-learning researchers interested in prompt-based (no weight update) alternatives that evolve repertoires of prompts.
Bridging KV-Cache Quantization and Linear Attention: From Theory to Pretrained Weight Migration
Who should care: Efficient-attention researchers: a theory linking KV-cache quantization to linear attention, plus migrating pretrained weights between them.
Lapras: Latent Reasoning for Time Series Language Models
Who should care: Time-series + LLM practitioners: replaces written chain-of-thought with latent reasoning for time-series question answering.
CARE: A Lightweight Plug-in Gated Correction and Uncertainty-aware Module for Long-term Time Series Forecasting
Who should care: Forecasting practitioners (energy, traffic, finance) wanting a plug-in correction and per-horizon uncertainty on existing models.
SDPAD: A Fully Spike-Driven Pipeline for End-to-End Autonomous Driving
Who should care: Energy-efficient autonomous driving: an end-to-end planner built entirely from spiking neural networks.
RAGNAROK: Radar-Aided Gravity-Normalized Alignment for Robust Open Keyframe-based Radar-Visual-Kinematic-Inertial SLAM
Who should care: Legged-robot state estimation in rough or dark conditions: radar-visual-inertial SLAM that limits yaw drift.
Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards
Who should care: Reasoning-model post-training: mixes RL with verifiable rewards and teacher guidance via a student-relative advantage.
Beyond Speech Captions: Speech-Rewarded Style Planning for Conversational Text-to-Speech
Who should care: Conversational text-to-speech: plans speaking style using speech-based rewards instead of text captions.
SAIL: Scientific Agentic Intelligence via a Science-Aware Loop
Who should care: People wanting an open science-research agent model (35B total / 3B active parameters) trained from diagnosed failures.
Toward Joint Optimization of Circuit Depth and Training Data Size in Adaptively Grown Quantum Classifiers
Who should care: Quantum ML researchers: trades circuit depth against training-data size in gate-by-gate grown quantum classifiers.
A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization
Who should care: Teams using LLM agents for expensive black-box optimization such as experiment design or tuning: a benchmark of agentic BBO.
Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits
Who should care: Tabular foundation model users (TabPFN-style) deciding which test-time compute tricks are worth the cost.
TACROSS: An Efficient and Low-Cost Scalable Human Touch System Across Heterogeneous Tactile Sensors for Dexterous Robot Learning
Who should care: Robot-learning teams collecting tactile demonstrations: maps cheap human tactile gloves onto robot tactile sensors.
CoPoE: Multimodal Fusion via Decomposable Disease-Coordinate Product-of-Experts for Missing-Modality Alzheimer's Diagnosis
Who should care: Clinical ML for Alzheimer's diagnosis when some data types (imaging, genomics, biomarkers) are missing per patient.
RaReCache: Bridging the Gap in Cross-Model KV Cache Reuse via Rank disagreement-based Selective Recomputation
Who should care: LLM serving teams that switch or cascade models mid-session and want to reuse the KV cache across models.
FloorSAV: Elucidating Spatial Audio-Visual Context with 2D Floormap for AV-LLMs
Who should care: Embodied AI with audio-visual LLMs: adds 2D floor-map context for spatial reasoning without fine-tuning.
SACQ: Structured Decoding with Memory-Conditioned Refinement for Long-Horizon Forecasting
Who should care: Long-horizon time-series forecasting: replaces the flatten readout head with structured, per-position decoding.
Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation
Who should care: Explainability researchers in vision: measures how well CNN, ViT and Vision Mamba attribution maps match human eye-tracking.
Kernelized Parameter Sensitivity Atlases for Amortized Conditional Interpretability
Who should care: Interpretability researchers needing per-input parameter attribution at scale: estimates sensitivities for new inputs without backprop.
Stochastic Teacher Intervention for Agentic On-Policy Distillation
Who should care: Teams distilling multi-turn agents: teacher interventions at random points to stop early student errors compounding.
Region-Aware CLS Token Augmentation for Fine-Grained Image Retrieval
Who should care: Fine-grained image retrieval: adds region-aware tokens alongside the ViT [CLS] descriptor.
MotherTree: Meta-learning on synthetic data improves decision tree training
Who should care: Tabular ML users who need auditable decision trees and want them pretrained on synthetic tasks for small datasets.
Symbolic Density Estimators for Unnormalized Distributions
Who should care: Scientists who want a closed-form probability density (an equation, not a neural net) from samples, including unnormalized cases.
Informationally Decoupled Trajectory Design for Sim-to-Real System Identification
Who should care: Robotics sim-to-real: designs test trajectories so each simulator parameter can be identified separately.
Strategic Governance of AI Models in Earth Science
Who should care: Earth-science and climate-AI governance: priorities for evaluating whether AI weather and climate models represent physics correctly.
EduBehaviors: Assertion-based Schemas for Auditable Coding of Educational Dialogues
Who should care: Education researchers coding classroom or tutoring transcripts: auditable LLM-assisted labeling of teacher talk moves.
Spiderbot: An Open-Source Energy-Efficient Hexapod with Passive Gravity Compensation
Who should care: Robotics hobbyists and multi-robot researchers: an open-source sub-$400 hexapod that uses springs to cut standing power by over 90%.
Relative Contact Velocity-Controlled Hand-Object Mechanism for Dexterous Tool Manipulation
Who should care: Dexterous-manipulation researchers: a hand-agnostic planning framework for picking up, repositioning and using tools.