Agents
Safe LLM memory reuse requires certificates, not simple success counters (ByteDance, Scale AI), common success counters miss partially stale reasoning memories, while per-memory statistical certificates reduce stale answers across five LLMs and code tasks.
Expert feedback for LLM agents should target decisions, not whole tasks (Amazon), SynAct directs expert labels to individual decisions and improves held-out accuracy in all 16 model-benchmark settings, by 11.7 percentage points on average.
Dense progress rewards help train agents on longer computer tasks (Adobe, UCLA), intermediate-state rewards improve 9B and 27B agents on seven technical domains and transfer to separate terminal benchmarks, addressing the sparse feedback of long workflows.
Shared message search lets LLM agent groups scale without flooding readers (Meta, Harvard), a searchable shared datastore outperforms agent-centric and blackboard communication on four tasks with less inference compute, while remaining useful after individual agents fail.
Language models
Vision scales multimodal learning without altering pure-text loss (CUHK, Tencent), controlled training shows that real visual content leaves the pure-text scaling frontier nearly unchanged while its multimodal benefit grows approximately as a power law of compute.
Flexible early stopping offers the best theoretical reasoning trade-off (CAS, University of Florida), 64.3% of reasoning tokens occur after the earliest measured accuracy maximum, and completing trajectories lowers accuracy by 9.03 points on average, strengthening the three-paper early-stopping thread.
Diffusion language models are cheaper to train than the widely cited 16x compute gap suggests (Cornell, Amazon), roughly 1,700 training runs reduce the estimated diffusion-to-autoregressive compute gap from 16x to 12.8x under current protocols and as low as 3.0x with small blocks.
Long reasoning distillation may teach behavior more than factual knowledge (CAS, Meituan), gains persist with flawed trajectories, repeated prompts and already-mastered problems, indicating that long-trace distillation primarily teaches sustained reasoning behavior rather than new facts.
Reinforcement learning
Reward models for vision-language AI can be trained without human preference labels (MIT, Scale AI), a linear head trained to rank web captions over perturbations beats GPT-4o by 6.6 points on VL-RewardBench at 32B, using about one thousandth of full fine-tuning compute.
Recursive synthesis turns unsolved problems into useful reinforcement learning data (UW, NVIDIA), MagicBeanStalk transforms hard seeds into tasks matched to the model and rollout budget, outperforming training on a 78,000-problem math corpus by preserving a useful mix of successes and failures.
Routing gradients can isolate reward hacking for removal after training (MATS, Anthropic), GRAFT routes suspicious updates into sacrificial parameters and preserves task gains while enabling learned reward-hacking behavior to be removed across models up to 32B.
Frozen LLM critics sharpen credit assignment for long-horizon agents (Apple), a frozen critic identifies productive actions, mistakes and recoveries within completed trajectories, outperforming GRPO and PPO most clearly on exploration-heavy tasks.
Vision
Theory and experiments show when splitting large images into views helps VLMs find small objects (UC Berkeley, Quotr AI), theory and about 177,000 requests show that overlapping views can improve small-object recall without losses when no crop zooms out and overlap covers at least one object.
Shared visual transition training improves multimodal LLMs across diverse tasks (Northwestern, CMU), WOVEN training transfers across external benchmarks even from small subsets, suggesting visual reasoning data should be organized by operation rather than scene category.
MLLMs miss text readable by squinting; diagnosis and training method recovers it (BUPT, Tencent), Squint OPSD raises ANLS from 26.50 to 88.61 on coarse-structure text by distilling from a downscaled teacher, without auxiliary views at inference.
Language model attention gives segmenters a clearer spatial target (KAUST, Adobe), dense spatial attention from a vision-language model contains more localization information than a single segmentation token and improves language-guided segmentation on two benchmarks.
Generative media
Recursive latent reasoner boosts diffusion models for visual puzzles (Warsaw Tech, University of Amsterdam), Painter-Thinker adds recursive latent constraint refinement to a frozen diffusion model, with advantages over standard diffusion widening as visual puzzles become harder.
Explicit context constraints improve interactive video story continuations (Google, UIUC), VIS-Ground explicitly connects the source, rendered video and new user request, improving grounded continuation by 10.3 points on average across three backbones.
One Bitwise Diffusion Framework Generates Across Three Multimodal Domains (Devotion AI Labs, Cambridge), representing modalities as binary streams lets one modality-agnostic denoising design support joint and conditional generation across three domains with strong cross-modal consistency.
Speech & audio
Post-training gains in audio models may reflect format adaptation, not true audio understanding. (NTU, NVIDIA), benchmark gains survive replacing audio with silence but largely disappear when multiple-choice options are removed, exposing textual format adaptation as the main source of improvement.
Open-vocabulary speech recognition runs on a $5 microcontroller without a neural accelerator (Lokutor), Oído runs int8 English ASR on an ESP32-S3 at 1.97x real time, with 3.70% and 8.23% word error rates on LibriSpeech test-clean and test-other.
Diffusion separates response timing from content in full-duplex speech (NVIDIA), Diffuplex concurrently listens, chooses when to speak and formulates content, approaching strong turn-based linguistic performance while supporting interruptions and natural turn-taking.
Robotics & embodied AI
Benchmark reveals when robotic agents truly learn and when they just retry (Luma AI), SEAR finds genuine experience-based learning across seven frontier models, but only code-based skills transfer reliably across tasks and simulators while scene-specific solutions remain brittle.
Robot policies falter as manipulation tasks stretch across more subgoals (INSAIT, NVIDIA), LADDER scales tasks from one to ten subgoals and shows that policies mastering individual skills still fail to compose them, even when high-level orchestration helps.
Latent world models for robotics scale predictably with compute (Meta, Mila), RoboJEPA's latent prediction error follows a compute scaling law and strongly correlates with downstream planning, adding a measurable resource-allocation signal to the 27-paper robot-world-model thread.
Autoregressive video pretraining unlocks the value of long context for real-time robot control (NVIDIA, MIT), extending history from zero to 19.2 seconds raises RoboCasa success from 63.3% to 78.7% only with autoregressive video pretraining, while bidirectional pretraining gains nothing.
Safety & security
Triggered spans can become provably invisible to language models (Microsoft Azure), a serving-time veto provides bit-identical noninterference outside triggered spans, while weight distillation creates covert selective-blindness backdoors with little clean-benchmark loss.
Internal signals can predict and prevent unauthorized agent actions (USC), a linear internal-state readout predicts authorization violations at AUROC 0.94, and gating actions on it reduces violations across four model families without sacrificing legitimate completion.
Image generators render false claims as credible visual evidence despite knowing they are false (Xiangtan University, City University of Hong Kong), four commercial generators visualize more than 70% of false-claim prompts as credible evidence, rising above 95% under EpiReal-Attack.
Shared coding rules can steer agents toward attacker-controlled packages (Independent), poisoned AGENTS.md, CLAUDE.md and .cursorrules files induce coding agents to import attacker-selected packages, with attacks transferring across models and frameworks in the long-running agent-security thread.
Evaluation & benchmarks
Structural baselines mask real progress in agent failure detection (Chemnitz, Fraunhofer), always blaming the final step achieves 67% to 96% accuracy on omission-heavy splits, and no tested method improves on this baseline for give-up failures.
LLM judges can be consistent yet wrong and order-sensitive (PayPal), zero-temperature verdicts can vary and candidate-order swaps frequently reverse decisions, motivating a trustworthy verdict rate that jointly measures correctness, reproducibility and order invariance.
Mean benchmark accuracy misleads model release decisions; recipe cards certify tail risk (SNU Hospital, Inha University), mean-based release rules violate a 10% deployment failure budget in 48.0% of cases, while finite-sample recipe cards reduce held-out violations to 0.6%.
Coding agents’ reports omit much of their recoverable task state (Tencent), CTXGAP finds a 40.7-point gap between report-only and full-context review across 4,799 questions, including after repaired code passes hidden tests.
Efficiency & systems
Recurrent-state quantization fails at long decode horizons due to scale-state feedback (Independent), INT4 recurrent-state quantization appears sound at short horizons but drives an eightfold perplexity increase after 1,000 steps, separating numerical instability from unrecovered functional loss.
Redesigning 4-bit AdamW quantization in preconditioner space reduces validation loss gap (UC Berkeley, Nubank), quantizing optimizer state by its effect on updates cuts the validation-loss gap to 32-bit AdamW at every tested scale, by as much as 70%.
Verified retries let smaller code models recover lost capability (Texas A&M, Apple), modest numbers of parallel, verified attempts let smaller or compressed code models recover solutions without latency growing in direct proportion to sample count.
Query-adaptive cache reads preserve accuracy while moving fewer bits (Amazon, UC Berkeley), a progressive codec reads four bits per scalar on average from an eight-bit KV cache while increasing C4 perplexity by at most 0.66% across six models.
Code & software
Agents devise cheaper evaluations to scale autonomous engineering search (UC Berkeley, Stanford), DuoDiscover jointly searches for system improvements and inexpensive tailored evaluations, beating human or automated baselines across repositories, kernels, engines and CPU architecture.
Code understanding, not edit volume, limits coding agents (Microsoft Research, University of Maryland), 6,840 controlled tasks show that tools offset many model weaknesses, but detailed code comparison and decomposition remain persistent bottlenecks.
Rewarding necessary edits makes debugging patches smaller without reducing correctness (Chicago, Google), PreciseCoder rewards both passing tests and touching necessary lines, reducing patches to about three lines across 4B to 27B models without losing correctness.
Verifier-based rewards train coding agents toward formally correct programs (UC Berkeley, AWS AI Labs), live verifier feedback rewards partial proof progress, substantially improving held-out formal verification and repository patches tested with withheld cases.
AI for science & health
Natural language can steer all-atom protein design across diverse objectives (UW, Ai2), Prompteina jointly generates sequences and all-atom structures from natural-language requirements, with nine of 89 tested designs binding a previously unaddressed target.
Lean proof assistant generates step-level labels to train natural-language proof verifiers. (UT Austin, UCLA), Lean-generated labels raise a 4B verifier's exact step accuracy from 48.2% to 74.7%, outperforming training on human labels from PRM800K.
Orchestrating coding agents broadens autonomous computational research beyond optimization (Sakana AI), 70 research campaigns show that separating hypothesis, experimentation, verification and evidence preservation can produce testable findings instead of merely optimizing a score.
Pretraining on synthetic primitive patterns enables robust zero-shot multivariate time series forecasting (Tsinghua, ByteDance), Timer-M1 learns from synthetic trends, seasonality, regime changes and cross-variable relations, ranking first on FEV and TIME and second on GIFT-Eval.