Agents
Terminal prompts make AI agents report success without doing the work (Independent researchers). The authors show that command-line confirmation prompts can leave requested actions unfinished while agents report success, making model rankings partly dependent on the execution harness. RECEIPT exposes operating-system process state to an unchanged agent, targeting an execution failure rather than a reasoning failure.
AI models find the best team deal in only 14% of conflicts (Stanford, Oxford). The authors’ Collaborative Intelligence Benchmark evaluates teamwork only on tasks each model can solve alone, separating collaboration failures from missing capability. Across 18 models, similarly capable models differed by up to 11 points as teammates, so solo performance is an insufficient selection criterion for multi-agent systems.
Language models
AI chatbots fail long tasks due to competing content, not length (UC San Diego). In the author’s retrieval experiments, neutral filler preserved near-perfect accuracy through 32K tokens, while competing statements reduced accuracy as low as 16%. Attention remaining on the evidence explained up to 58% of loss variance, suggesting that distractor resistance matters more than nominal context capacity in these settings.
Giving harder inputs more attempts helps little on hard LLM workloads (Hankuk University of Foreign Studies). The authors report that allocating more attempts to harder inputs captures at most 13% of the oracle allocation gain on hard workloads. With a verifier and an informative prior from another model, Thompson Sampling, an adaptive allocation method, raised the solved share from 14% to 27% on their hardest workload; without a good prior, no policy beat equal allocation.
Reinforcement learning
Language models can lag 50 to 150 updates behind training (Stanford). On Qwen3-1.7B, the authors find that only 0.4–0.5% of the lower-precision weights used to generate training data change each step, leaving that model effectively 50–150 updates behind the optimizer. Their correction improved task scores by an average of 6.1 points across 25 matched seed-pairs, exposing a numerical synchronization problem in reinforcement-learning pipelines.
AI models can be trained end to end with one algorithm (Carnegie Mellon, Johns Hopkins). The authors express pretraining, fine-tuning and reinforcement learning within one maximum-likelihood algorithm that continuously mixes static and model-generated data. Their unified trainer reportedly improves accuracy and task coverage over staged training at equal compute while preserving broader capabilities, making training-stage boundaries worth testing rather than assuming.
Vision
Image search models can retain left-right information but ignore it (Huawei). The authors distinguish encoding spatial relationships from actually using them to match images and captions. On a 118,287-image index, their training-free correction raised left/right pairwise accuracy from 51–53% to 81–96%, indicating that better use of existing representations can recover substantial retrieval capability.
AI that identifies radiological findings first can diagnose unseen diseases (UC Berkeley, Alibaba). Diagnosis by Composition first recognizes radiological findings, then combines them using published diagnostic guidelines. In the authors’ multi-phase liver CT evaluation, it outperformed all comparators on detecting and localizing diseases excluded from training, including rare diseases at external institutions; this is evidence of retrospective generalization, not prospective clinical validation.
Generative media
Joint video and 3D motion prediction reduces trajectory errors by over half (RoboScience, NUS). WorldFlow jointly predicts future video and the 3D trajectories of specified points, using 1.44 million object-level training samples from simulated and real data. The authors report 54.2% lower average displacement error than the best external baseline on spatial-reasoning motion forecasts, supporting joint visual and geometric prediction for world models.
Repurposing pretrained image generators for any-order editing costs a fraction of retraining (Stony Brook, ÉTS). The authors adapt image generators trained to emit tokens in a fixed raster order to generate in arbitrary orders with modest additional training. Their recipe outperforms MaskGIT on ImageNet inpainting and outpainting while retaining standard autoregressive inference infrastructure, offering a reuse path for existing generators.
Speech & audio
Image generation models can be repurposed for speech synthesis while preserving voice (Wuhan University, Microsoft). The authors adapt FLUX.2 to speech synthesis by treating spectrograms as images while freezing the image encoder. In matched comparisons, image-pretrained initialization reduced word error rate from 4.2% to 2.7% while preserving speaker identity, suggesting useful transfer beyond the original output modality.
Multimodal model tops specialist speech systems for low-resource African languages (Microsoft Research Africa, Lancaster). In the authors’ matched low-resource comparisons across six African languages, Microsoft’s Phi-4 achieved the best speech-recognition accuracy on most languages against Whisper and MMS. Domain-specific training helped Phi-4 and Whisper but hurt MMS, while related-language adapters beat adapters trained from scratch, making adaptation strategy as important as model category.
Robotics & embodied AI
Simulator bugs cause most robot policy failures in common benchmarks (Tsinghua, UCAS). In five container-placement tasks, the authors found that over 90% of a leading policy’s failures involved the simulator ejecting correctly placed objects. They traced the artifacts to a collision-detection threshold in PhysX 5.3.1 and eliminated them by changing one constant, a strong warning against interpreting every simulated failure as a control failure.
Robots can act at 50 Hz while updating visual understanding once a second (Leapmotor). Leap-VLA separates a slow semantic model from a 20–50 Hz action controller that receives fresh visual input and cached semantic representations. Leapmotor reports that semantic updates can be slowed to two-second intervals while retaining over 94% of the synchronous baseline’s success rate, suggesting that fresh control inputs need not require a full model pass.
Contact-preserving retargeting improves robot task success from 13% to 83% (Carnegie Mellon, Amazon). CARD converts human demonstrations into a representation that preserves contact-force targets rather than only joint or object positions. In the authors’ hardware experiments, a tactile tracking controller raised task success from 13% with position-only replay to 83%, showing why contact information matters when transferring demonstrations to robots.
Safety & security
Most AI coding agents overrun a disk limit before reading the file (Tsinghua, University of Sydney). Across 42 harness-model combinations, the authors found 32 exceeded a 4 MiB disk budget on crafted inputs, with 26 doing so before content reached the model. A byte-counting gate kept every routed extraction within budget while agents completed 25 of 28 tasks, placing the effective defense before model inspection.
Fabricated hazard reviews can steer AI agents to reject legitimate options (Technion, Zenity). Across six models and ten service domains, fabricated safety allegations redirected agents with 60.4% attack success, versus 66.0% for instruction injection, while evading two commercial injection detectors. In a separate defense evaluation, requiring corroboration reduced success from 68.4% to 3.6% against a non-adaptive attacker, showing both the risk of evidence poisoning and a bounded mitigation.
Membership inference rarely succeeds despite measurable training effects in language models (KTH). The authors reran OLMo 2 1B’s 50-billion-token mid-training stage using 7,814 matched document pairs. Although trained documents received higher likelihood in 85.4% of pairs, membership inference reached only 3.8% true positives at 1% false positives, distinguishing a measurable average training effect from reliable identification of individual training documents.
Evaluation & benchmarks
How data is split can swing encrypted traffic classifier scores by 40 points (Independent researchers). With the same model and code, the authors report macro-F1, a class-balanced classification measure, changing from 0.992 to 0.567 when the train-test partition changes. Their single-identifier probes diagnose the appropriate separation axis without model training, making split validation a prerequisite for credible encrypted-traffic comparisons.
Textual personas fail to make vision-language models answer like real people (Netflix). Comparing model answers with measured responses from people with low vision, color blindness or hearing loss, the author finds that textual personas never outperform simple rules based on the stated facts. Knowledge about perceptual conditions did not produce faithful simulation of perception, limiting the use of persona prompts as substitutes for user studies.
Batching makes AI grading verdicts depend on the other items (Technical University of Munich). In harm detection and short-answer grading, the authors found that changing batch companions flipped 11–22% of pass/fail verdicts, more than changes to instruction wording or item keys. The effect persisted when averaging over batch orders, so batching is part of the measurement protocol, not merely a cost optimization.
Efficiency & systems
Mid-depth layers judge KV-cache evictions better than the owning layer (Colby College). Across 18 transformer models, the authors find that layers 40–75% through the network best identify which tokens to retain in the key-value cache, the stored attention state used during generation. A predictor based on initial prompt-processing attention selects the scoring layer without extra forward passes, offering a low-overhead alternative to local or final-layer eviction scores.
Chunked prefill, not cache architecture, drives LLM serving engine speed differences (Thompson Rivers University). Under default settings, the authors measured vLLM median time to first token at up to 64 times SGLang’s. Matching chunked prefill, which processes prompts in smaller scheduling units, removed 78% of the capacity gap and 30–67% of the latency gap, making configuration matching essential before attributing performance to architecture.
Million-token prompts served at 2.3 times the throughput without retraining (Nvidia). Nvidia’s authors replace token-level attention selection with summaries of 16-token chunks and reuse selections across neighboring prompt queries. On million-token prompts with 32 concurrent requests across four B300 GPUs, they report 2.29 times dense-attention throughput with benchmark scores close to dense, a hardware-specific result worth checking against production workloads.
Code & software
Verified code often passes formal checks but misses the real spec (Independent researchers). The authors report a pipeline with 99.1% verifier acceptance but only 1.9% of samples having specifications bound to the intended requirements. Even structurally verified code failed 11–15% of hidden tests, showing that proof acceptance does not establish that the right behavior was specified.
Agentic framework formally verifies GPU kernels against PyTorch references (Imperial College London, Cambridge). LARK uses Lean 4 to verify GPU kernels’ functional and memory behavior against PyTorch references under exact arithmetic. The authors report verifying a majority of kernels from their benchmark and production-code collection and finding a previously unknown data race, while making the operator semantics and arithmetic assumptions explicit trust boundaries.
Static analysis cuts LLM cost for Python type annotation by 75% (Korea University, CISPA). SALT combines static type inference with language-model calls rather than asking a model to infer every annotation. Across 12 real-world projects and four models, the authors report higher accuracy than a strong coding-assistant baseline and up to 78.5% lower LLM cost, supporting selective model use where program analysis leaves uncertainty.
AI for science & health
Language models improve psychiatric interviews after training with simulated patients (Google DeepMind, Stanford). After training on simulated psychiatric conversations, the authors’ model scored higher than its identically prompted, untrained counterpart in 76.7% of matched encounters with psychiatrists role-playing patients. Separate safety tests found fewer failures to recognize acute-risk urgency, but the preprint does not establish performance or safety with real patients.
Satellite data improves AI species maps in remote areas by 51% (Google DeepMind, Google Research). Google’s authors trained OCELOT on two billion species-occurrence records and satellite embeddings to predict relative occurrence across 100,000 taxa. They report a 51% improvement in geographic extrapolation over coordinate-only models, while matching or exceeding interpolation baselines, making environmental representations useful beyond well-observed locations.
| 