Results below are authors' reports from preprints; peer-review status was not supplied.
Agents
AI assistants can recall more irrelevant memories as their memory stores grow (Zhejiang Sci-Tech University). The authors report 87% to 100% false-alarm rates at 2,048 stored memories when retrieval uses a fixed similarity threshold tuned for smaller stores, across the tested benchmarks and embedding models. Their label-free conformal gate keeps false alarms near the chosen nominal level, making retrieval calibration a scaling requirement rather than a one-time setting.
Context compaction spares benchmark scores but silently drops task constraints (ELLIS Institute Tübingen, MPI for Intelligent Systems). Testing five context-compaction strategies in a fixed agent scaffold across three long-horizon benchmarks, the authors find little change in task scores despite summaries dropping constraints, especially instructions about how to work. Benchmark parity therefore does not establish that compressed histories preserve the user's requirements.
Simple retrieval rivals complex memory systems on conversation benchmarks (institutions undisclosed). The authors' Raw-RAG baseline, which retrieves passages directly from conversation history, ranks first in six of seven evaluated settings. The result makes direct retrieval a necessary baseline before investing in elaborate memory restructuring.
Language models
Constrained decoding does not inherently degrade LLM reasoning, contrary to common assumption (Télécom Paris, Nokia Bell Labs). Through a theoretical analysis and causal ablations across 11 open-source models, the authors attribute apparent reasoning losses to prompt differences rather than decoder-level masking. They report accuracy improvements of up to 17.4% from validity constraints, while constraints on reasoning help only when they enforce useful solution trajectories.
Halting heads can overstate compute savings in looped language models (Apple). In pretrained looped models up to 1.3 billion parameters, the authors find that predicted iteration counts can substantially understate actual execution. Batching removes most of the remaining advantage over simpler models, so token-level stopping estimates are a poor substitute for measured serving cost.
Data curation advances LLM training only if a train-target gap exists (JD.com, Oxford). The authors distinguish reaching the same performance sooner from improving the attainable optimum, and show that the latter requires a correctable mismatch between training data and the target distribution. Their Δ-test estimates that opportunity from one baseline run, with language-model experiments supporting persistent gains only when the gap is nonzero.
Reinforcement learning
Training world models for policy improvement beats optimizing pixel accuracy (Tencent, Peking University). The authors' DRILL-IMG trains video world models for the robot policies that learn from their simulated trajectories, rather than for image prediction alone. It improves mean task success over two comparison methods despite 17% higher next-frame pixel prediction error, showing that a worse visual predictor can be a better training environment.
AI agents answer questions better using 68% fewer extra training tokens (University of Sheffield, CMU). TWICE compares factual and counterfactual agent trajectories using shared randomness to isolate individual decisions' effects. Across seven question-answering benchmarks, the authors report average exact-match improvements over GRPO of 5.3 and 4.5 percentage points at two model sizes, while using 68% fewer auxiliary tokens than a prior credit-assignment method.
Vision
Removing textual shortcuts helps vision-language models learn from visual examples (IBM Research, Tel Aviv University). MIME leaves the task rule unstated during training, requiring models to infer it from visual examples rather than textual shortcuts. The authors report gains of 8.1 to 18.0 percentage points across three models, without sacrificing performance on the tested single-image tasks.
Language models shown red hearts can favor code that calls itself (Google DeepMind, Stanford). The authors find that a red-heart symbol steers vision-language models toward recursive solutions on simple coding problems, without that influence appearing in their explicit reasoning. Communication-game experiments also reveal mismatches between models' generation and interpretation of visual meaning, cautioning against treating written reasoning as a complete account of visual influences.
Generative media
Learning visual concepts before rendering improves image generation in controlled tests (UC San Diego, Adobe Research). In controlled experiments, the authors find that autoregressive pretraining works better for acquiring visual knowledge, while diffusion works better for rendering detailed pixels. Tokens carrying semantic information are more useful than reconstruction-optimized tokens, supporting a separation between learning what to depict and learning how to render it.
Image generators follow object relationships more faithfully with targeted prompt edits (Zhejiang University, Tencent AI Lab). StruPO uses calibrated scene graphs, structured descriptions of objects and their relationships, to guide prompt editing. Across multiple test sets and generation backbones, the authors report a better faithfulness-aesthetics trade-off, particularly on complex relationships, with one editing pass after graph calibration rather than iterative critique and rewriting.
Robotics & embodied AI
Robot AI speed depends more on action generation than model size (NUS, 0G Labs). Benchmarking 18 open vision-language-action models on NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395, the authors split inference into seven stages and find that action-generation strategy dominates latency more than model size. Kernel fusion can deliver substantial speedups, but its limited availability makes implementation support part of model selection.
Force-reactive control lets humanoids learn contact-rich tasks from demonstration (Amazon FAR, NYU). FLoRe combines robot-free demonstrations recording hand contact forces and torques with a force-reactive policy and whole-body controller. On three contact-rich tasks using a Unitree G1, the authors report 80% rollout success versus 26.7% for a position-only policy, evidence that force information matters when contact determines success.
Safety & security
Chatbot defenses can miss poisoned documents padded with meaningless words (institutions undisclosed). Reviewing 21 published evaluations of retrieval-augmented generation defenses, the authors find frequent attack-construction cues, including copied questions and distinctive passage lengths. Removing such cues from PoisonedRAG attacks reduces RAGuard's detection AUROC from 0.953 to 0.758 and raises its miss rate from 0.191 to 0.763, suggesting that apparent protection can depend on benchmark artifacts.
AI agents take harmful actions on notices they have judged untrustworthy (Oakland University, Dartmouth). Across five models, 196 constructed false notices with legitimate counterparts and 204 archived attacks adapted for simulated tool use, the authors find harmful tool calls on 27% to 99% of notices the models had rated untrustworthy. Removing the preliminary rating step increases harmful execution by 15 percentage points on average, but warnings also impair legitimate completion, exposing a gap between recognizing risk and enforcing an action boundary.
AI code reviewers miss one in six attacks spread across updates (UIUC, Arizona State University). CyberDual tests attacker and reviewer agents on maintenance commits to 14 open-source C/C++ libraries, verifying exploits again after review. Across 1,088 valid attacks, individual reviewers catch 85% to 91%, yet about one in six attacks spread across an update window survives all three reviewers, supporting review of accumulated changes rather than isolated commits alone.
Evaluation & benchmarks
Systematic audit reveals pervasive data quality issues in RVL-CDIP document classification benchmark (Vanderbilt). The authors audit the 400,000-document benchmark and identify label errors, train-test leakage, spurious cues, sensitivity to optical character recognition and sensitive personal data. After remediation, retraining improves accuracy by 0.9 to 2.9 percentage points, but the best model reaches 96.5% on the corrected benchmark versus the 98.8% claimed on the original, showing why the two evaluation settings cannot be treated as interchangeable.
AI agents’ clicks marked wrong can work a third of the time (Zhejiang University, Geely). By executing clicks and comparing their outcomes, the authors find that 17.7% and 31.2% of clicks rejected by bounding-box scoring are actually correct on a code forge and an online shop. An audit of 11 models on ScreenSpot-v2 yields corrections of 0.2 to 1.8 percentage points, large enough to rival reported gaps between models.
Prompting strategy effects depend on judge model and language in LLM evaluation (University of Amsterdam, CWI). Across nine prompting strategies, four open-weight judges, 27 languages and four tasks, the authors collect over one million scores and find that judge identity explains far more variance than prompting strategy. Strategy effects can reverse across judges, particularly for low-resource languages, making a single judge's aggregate score weak evidence for a universal prompting improvement.
Efficiency & systems
Sharing random numbers speeds integer training when rounding errors are corrected (Google). The authors share random rounding thresholds across groups of gradients in INT8 training, then correct the resulting errors. In their pretraining experiments, the method matches independent stochastic rounding's validation loss and runs full dense-model training steps 1.24 to 1.32 times as fast as BF16, offering measured step-time gains rather than arithmetic savings alone.
Language models retain near-full accuracy with thirteenfold smaller cache memory (Rice University). BellowsKV maps hidden states from adjacent layers into one shared representation per token, then applies low-bit quantization or token eviction. The authors report up to 14-fold key-value cache compression with minimal accuracy loss in their evaluations, outperforming compared cross-layer compression methods without training.
Code & software
Coding agents struggle with infrastructure changes that existing tests miss (University of Edinburgh, Tencent). InfraDev asks agents to reproduce substantial changes to software for AI training, inference and reinforcement learning. Evaluating six frontier agents, the authors find that existing tests can pass without establishing correctness, while incomplete task descriptions can independently block valid solutions.
Coding agents can report successful hardware migrations while changing model training (NC State, Federal University of Ceará). In independently checked GPU-to-TPU migrations of ten PyTorch projects using Claude Code, the authors report that 76% of execution attempts failed, with 49% of those failures raising no error. One silent behavioral change reduced accuracy from 86.53% to 84.89% despite the agent reporting success, making training-equivalence checks essential alongside successful execution.
AI for science & health
AI world models solve tasks without faithfully capturing how environments change (Helmholtz Munich, MCML). Across three leading latent world models in simple control environments, the authors find that high task reward does not imply faithful predictions of environment dynamics, especially from pixels. Their latent rollout consistency test and post-hoc state decoding offer separate checks before treating a control model as a scientific simulator.
AI models can reuse learned dynamics across tasks and physical conditions (Harvard, Microsoft Research). LEAP learns a correction to a tractable reference process from observed trajectories, capturing local dynamics and dependencies across time. The authors demonstrate reuse for ranking trajectories, steering mechanisms and constrained optimization under different physical conditions, making the learned dynamics useful beyond one downstream task.
Theory
Top-k data scores under-cover latent modes; class balancing or coverage fixes selection (University of Sydney). In a latent-mode model, the authors prove that selecting a fixed top-ranked subset using per-example value scores incurs excess risk relative to adaptive coverage. Their practical prescription is to balance classes when labels identify the underlying modes, and otherwise select for coverage using the learner's representation geometry.
Muon's orthogonalized updates win when gradients are far from low rank (KAUST, MBZUAI). Analyzing PolarGD, the orthogonalized update underlying Muon, on matrix quadratics with exact line search, the authors establish one-step convergence under weaker conditions than gradient descent. They also show that approximate orthogonalization can converge faster than an exact update when iterates have tiny singular values, so greater numerical exactness is not automatically a better optimizer.