News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 43 of 54 · The Research front page
Zhang et al. introduce CypherTurn, a benchmark for generating Cypher graph queries across conversations in which later requests depend on earlier results. Testing 15 models under guided and autonomous protocols, they find that turn-level competence does not translate into reliable sessions, and that managing errors is partly distinct from generating correct queries.
Tao et al. audit multi-hop question-answering benchmarks by removing evidence for individual reasoning steps and checking whether models can still recover answers they previously got right. Existing benchmarks show weak dependence on their annotated chains, while the authors' REALHOP framework raises that dependence substantially and better separates model performance.
Wang et al. present BlindBias, a jailbreak framework for black-box language models that expose sampled continuations but not internal probabilities. Across four target endpoints and three safety benchmarks, the method generally achieved higher mean evaluation scores than the tested baselines while reducing queries through selective intervention and speculative execution.
Chen et al. introduce Hindsight-Divergence Localization (HDL), a method for reinforcement learning with verifiable rewards that concentrates exploration where feedback causes a model to reconsider earlier choices. Across three models and tasks spanning mathematics, code, and agents, HDL generated up to 2.5 times fewer tokens than GRPO while improving downstream performance.
Gong and colleagues isolate the effect of message passing in multi-agent LLM systems by holding a downstream agent’s task and evidence fixed while varying the upstream message. They find that communication can correct an initially wrong receiver, but an incorrect message can overturn an initially correct answer in up to 32% of cases, usually by inducing the message’s specific error.
Qi and colleagues introduce VACE, a training loop that alternates updates to a language model’s weights with revisions to its harness, the instructions, skills and execution rules surrounding the model. Using Qwen3.5-9B, they report that accepting harness changes only after validation produced better results than weight-only reinforcement learning or alternation without a validation gate.
The authors introduce Grounded Entity Biographies (GEB), a long-video memory framework that groups visually grounded observations of the same physical instance across clips into retrievable biographies. On four benchmarks, including day-long and week-long recordings, GEB outperforms prior memory frameworks in both multiple-choice and open-ended question answering. On EgoLifeQA it reaches 72.0%, 4.4 percentage points above the best published result.
Liang et al. propose πPPO, an actor-critic framework for reinforcement learning with verifiable rewards (RLVR) that conditions the critic on contrastive evidence from successful and failed rollouts of the same prompt. This self-privileged formulation consistently improves value-estimation accuracy and outperforms standard PPO and critic-free baselines on challenging mathematical reasoning benchmarks, while also supporting smaller asymmetric critics.
Ying et al. introduce Exemplar2VQA, a multi-agent framework that synthesizes large-scale 3D spatial question-answer pairs in simulated environments. By converting user-defined exemplar queries into deterministic code via specialized agents, the system avoids hallucinations common to LLM-based generation. Fine-tuning Qwen2.5-VL models on this synthetic indoor data improves performance on multiple spatial benchmarks, including outdoor and mixed-scene settings.
The authors propose GC-OPD, a distillation method that gives an off-the-shelf teacher access to a graph of its own past executions when scoring student trajectories. Across ScienceWorld, ALFWorld Unseen and WebShop, it improved mean success over vanilla on-policy distillation, and on ScienceWorld and ALFWorld it also surpassed OPD baselines using GRPO-trained teachers, without retraining the teacher.
Tirosh, Amos and Geva introduce LIFT, a Transformer architecture that feeds a compact continuous state from each generated position into the processing of the next. Across models from 135 million to 1 billion parameters, the authors report better language modeling, reasoning and procedural-task performance than standard Transformers trained on the same number of tokens, with competitive or better results when training compute is matched.
Oota et al. separate agent planning into two problems: choosing an appropriate planning mode and executing that mode faithfully. Across four benchmarks and three language models, generic Plan+ReAct agents frequently abandoned their declared structure, while pattern-specific executors largely closed this execution gap. The remaining weakness was selection, since models did not reliably choose the best mode for each task.
Okonkwo et al. analyze why group relative policy optimization wastes computation on prompts whose sampled answers are either all correct or all incorrect, producing no learning signal. They derive a prompt curriculum in Fisher-Rao arc-length coordinates and implement it in ARCUS, which outperformed standard GRPO and dynamic sampling across six mathematical reasoning benchmarks and three model backbones.
Qi and colleagues examined how maintainers handle performance fixes submitted by coding agents, then reran a subset of accepted and rejected patches. Although 57% of closed fixes were merged, only 18 of 30 re-executed merged fixes met the authors’ delivery criterion, showing that acceptance is not reliable evidence of a real performance improvement.
Tang, Liu and Sarabi address a central problem in persistent language-model memory: a rewrite may help only much later, and its apparent value may actually come from information already stored. Their Memory Gain Policy Optimization method compares updated and prior memory states to isolate the value created by each rewrite, improving document-level information extraction while cutting average memory length by nearly 80% from the initial policy.
Paulo et al. examine why fine-tuning on narrowly harmful advice can produce unrelated harmful behavior, a phenomenon called emergent misalignment. Using training data attribution and controlled retraining, they show that harmful examples contribute very unequally: small, carefully selected subsets can produce sharply different misalignment rates.
Dixit et al. introduce FOCUS, a training-free layer that removes parts of an agent’s interaction history according to their estimated influence on future decisions. Across tool use, question answering, web navigation and multi-turn dialogue benchmarks, the authors report lower context requirements and, in some settings, better task success than retaining the full history.
Mao et al. present TIDE, a framework that evolves an agent’s external memory using delayed signals such as clicks, conversions, and negative feedback while leaving the underlying generative model fixed. In e-commerce content generation, the approach improved performance in temporal replay and an online A/B test, while also outperforming comparison methods on a delayed-label benchmark.
Shao and colleagues introduce Context Language Models, or CLMs, which let a language model directly rewrite the information it retains rather than relying on an external harness to summarize or discard context. Across several long-horizon benchmarks, existing models using this design outperformed engineered context-management strategies while consuming fewer floating-point operations, or FLOPs.
Chu et al. introduce Retrieval-Augmented Skill Optimization, or RASO, for improving the natural-language instructions that guide agents inside particular tool and execution environments. Across four benchmarks and two language models, the authors report that retrieving and adapting existing skills consistently outperformed optimization baselines that lacked retrieval during initialization and updating.
Kirchhof et al. introduce RLTL;DR, a training method that turns verifier feedback from failed coding and tool-use attempts into short, model-generated insights for subsequent attempts. On tasks the base model could not solve in 128 tries, the method raised Pass@1 from 0 to 1 percent under standard training to as high as 31 percent, while retaining 12 to 13 percent when the insights were removed at evaluation.
The authors present PRISM, a real-to-sim-to-real pipeline for training humanoid robots from a small number of human interaction videos. By generating varied counterfactual videos and grounding their reconstruction in human-object contact, they train one depth-based policy that picks up, carries, and drops novel objects on a real robot without real-world fine-tuning.
Wang et al. from Alibaba AAIG propose InGuard, a safety framework that operates inside the T2I pipeline using the model's own text encoder embeddings and intermediate latent states. Across five open-weight T2I models, InGuard achieves 97.9-98.8% safety rate while disturbing benign prompts 57.5-73.5% less and skipping 50-55.6% of denoising steps compared to conventional outer guardrails.
The authors introduce RECAP, a redundancy-aware credit assignment method that assigns step-level rewards based on both a step's downstream role in the reasoning structure and its progress toward the correct answer. Applied to two 7B models across four math benchmarks, RECAP improves pass@1 while cutting reasoning tokens by 8%-31% relative to GRPO.
Li et al. train an agent harness from a controller-free scaffold using execution failures, localized code repairs and held-out validation. Across two web-agent benchmarks and models ranging from 4B to 120B parameters, the resulting harness matched or exceeded competing methods in most settings while cutting model calls by up to 91.8%.
Pan et al. introduce a model for continuous interaction with streaming time-series data, rather than waiting for a complete sequence or pausing input while responding. On their StreamTSI-34K benchmark, TimeInteract outperformed language, vision-language, and time-series language models while nearly eliminating stream stalls.
Shen et al. generated 3,300 stateful agent environments by first sampling and solving mathematical mechanisms, then translating them into tool-based tasks with hidden or revealed parameters. Training a 35-billion-parameter Qwen model on three mechanism families substantially improved interactive performance and transferred to unseen families and external agent benchmarks.
Zhang et al. extend on-policy self-distillation from selecting interface elements to completing multi-turn GUI tasks. Their two-stage method first teaches the model to use corrective guidance, then distills step-specific reasoning and task-relevant memories into the agent, improving success rates on AndroidWorld and MobileWorld.
Qadri and colleagues studied an autonomous agent embedded across several teams at a large technology company, using semi-structured interviews to examine its role in everyday collaborative work. They report that adoption depended not only on technical capability, but also on whether the agent respected unwritten communication norms and allowed people to retain meaningful control.
The authors propose concurrent confidence calibration, where confidence is learned alongside capability improvement during reinforcement learning from verifiable rewards (RLVR). Their method, CoCal (Companion Confidence Calibration), trains a separate lightweight model on rollout hidden states and verifier feedback, decoupling confidence estimation from policy optimization. Experiments on Qwen3-8B and Qwen3-14B show that CoCal improves confidence estimation without sacrificing task accuracy, outperforming both RL-based concurrent methods and post-hoc calibration.
The authors propose Decoupling and Coupling Reinforcement Learning (DCRL), which reframes LLM reasoning as a coupled geometric manifold and aligns policy and reward models during training. In experiments, a Qwen3-4B model trained with DCRL surpassed a Qwen3-32B baseline and approached the performance of a Qwen3-235B model.
The authors analyzed 1,248 Markdown-based GitHub Agentic Workflow files across 276 repositories. They found that workflows contain a median of 556.5 words and are maintained for months, yet only 9.4% address prompt-injection attacks. The study also created a taxonomy of instruction types and evaluated LLM-based labeling.
The authors propose a method where a small concept generator is trained using reinforcement learning to produce high-level concepts (e.g., hints, strategies) that maximize the downstream success of a frozen, larger answer generator. On hard mathematical reasoning problems, this learned search policy substantially improves pass@k over naive repeated sampling at the same answer generation allocation, transfers to other answer generators, and outperforms concepts from much larger untuned models.
Li presents an incremental compiler that converts written agent skills into extended finite state machines, which track data and execution progress while enforcing transitions between operations. Across four benchmarks and four executors, HEXIS improved success over a Skill + ReAct baseline by an average of 16.2 percentage points.
Wang built an agentic robotics system that diagnoses failures, writes skills or installs external models, and tests changes without human-written robot code. The agent successfully discovered some missing capabilities, but its improvements did not accumulate into task success because relational perception, sequential skill testing and the evaluation harness constrained what it could learn.
Ding et al. introduce 136 deployment tasks evaluated through programmatic checks in freshly rebuilt environments, without an LLM judge. Seven models resolved 52.9 to 75.0 percent of tasks, with substantially stronger results on repairing existing configurations than on creating new ones.
Xu et al. propose Constraint-Driven Context Engineering, a method for turning domain constraints into structured interfaces between AI systems and their operating environments. Applied to educational assessment, healthcare decision support and financial-distress prediction, the method distinguishes constraints that guide model behaviour, enforce boundaries or help verify outputs.
Gan et al. built WebMRE from successful WebArena trajectories to study web-agent training without the variability of live websites or model-based judges. Their experiments suggest that generating a human-readable guide can help a model choose the corresponding interface action, although the gains from joint training were small and depended on model scale and output order.
Stein et al. convert planning domains from PDDL into Lean, then use GPT-5.6-Sol to generate both a generalized plan and a formal proof of its completeness. Lean's kernel accepted proofs for 12 of 13 benchmark domains, replacing finite test coverage or manual inspection with machine-checked guarantees relative to the formal specification.
Xu et al. address a mismatch in disaggregated LLM serving, where separate pools handle prompt processing, called prefill, and token generation, called decode, even though demand for each phase changes rapidly. Across public and internal traffic traces, Crossflow increased token throughput by 16.2 to 17.4% on geometric mean over static partitioning, while reducing average time to first token at every evaluated load.
Wu et al. developed Env-Rethink, a framework built around a post-trained 27-billion-parameter model that prepares persistent file environments for LLM agents. Across 30 tasks and nine models, the authors report a 15.1 percentage-point increase in mean rubric pass rate after environment preparation.
Jiao et al. introduce an online reinforcement-learning method that assigns counterfactual credit to a model's choice of reasoning mode, while training its answer tokens with a separate objective. Across nine benchmarks and two Qwen3 scales, the method improved macro-average accuracy while reducing generated tokens by 51% and 41% relative to always-thinking checkpoints.
Ma and colleagues train agents using changes in verified task progress rather than relying only on final success or failure. Across three Qwen3.5 model scales on AppWorld, ProCredit achieved higher task completion rates than outcome-reward and progress-based baselines, including a 4.1 percentage-point gain over the strongest outcome-reward baseline at 4B.
The authors introduce a form of on-policy distillation designed for prompts containing many simultaneous constraints. Across two Qwen teacher-student pairs and seven benchmarks, their method achieved the highest average performance among the evaluated training approaches, and a 1.5B student surpassed its 7B teacher on MulDimIF.
Wu et al. replace independently sampled reasoning attempts with a planner that jointly assigns different solution outlines before a fixed model executes each one. Across five mathematical reasoning benchmarks, their zero-shot method improved pass@64 by up to 6.7 percentage points over repeated sampling, while a reinforcement-learned planner improved it by up to 13.4 points.
Meng et al. built a long-term agent memory architecture that partitions interactions into thematic episodes and indexes facts by entity, type, property, value, time, and provenance. On the LoCoMo and LongMemEval benchmarks, the authors report high answer accuracy with smaller, evidence-focused contexts rather than full histories or lossy summaries.
Zhao, Yu and Zhao study stable-but-wrong lock-in, where a diffusion language model settles early on an incorrect answer despite having denoising steps left. Their LOCKR planner uses the trajectory of internal hidden states to trigger additional computation and choose repairs, improving mathematical reasoning accuracy by 2.21 to 5.37 percentage points across five settings.
AlRabah et al. developed a memory layer that lets EDA agents reuse procedures only after the resulting chip design artifacts pass synthesis, simulation, or formal checks. Under matched model and tool settings, the system improved equivalence-passing RTL rewrites and transferred a frozen skill library to unseen verification tasks.