News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 48 of 54 · The Research front page
Shen et al. target the branch-resolution barrier, where an agent must wait for a model output or user choice before beginning downstream work. Across four agentic workloads, DynBranch reduced mean latency by up to 32% relative to each workload's strongest prior system while preserving workflow results.
The authors built an automated white-box auditing system that analyzes an AI agent's repository for attacker-controlled paths to sensitive operations, then tests concrete exploits in a controlled runtime. Across 72 reproducible vulnerabilities in 12 open-source systems and frameworks, AgentXploit achieved 59.3% end-to-end success over three runs.
Ahmed, Poech and Röttger tested retrieval-augmented generation on 317 documented knowledge reversals spanning medicine, law, software and platform policy. They found that stale documents can turn otherwise-correct answers into wrong ones, while explicit applicability information and recency-aware retrieval reduce the failure.
The authors built an extensible evaluation platform in which language models compete directly in structured games rather than answer a fixed test set. Its three pilot environments test strategic planning under perfect information, hidden information and multiplayer social interaction, with wins, losses and draws providing ground-truth evaluation.
Hosseini et al. trained reasoning models to estimate confidence at intermediate points in their own reasoning, using labels derived from the models' token probabilities and only 600 training problems. The resulting models generated up to 25% fewer tokens at matched accuracy, despite receiving no objective to reason briefly or stop early.
Du et al. address a trade-off in representation autoencoders: shallow visual features preserve pixel detail, while deeper features can support better image generation. On ImageNet-256, their randomized layer-fusion training improves reconstruction flexibility and reduces unguided generative FID by up to 29%, without modifying the pretrained DINOv3-L encoder.
Xia et al. train a model to handle long inputs in three explicit stages: identify question-relevant evidence, combine it into a compact summary, then answer. On a seven-task evaluation suite, H2S-14B achieved an average score of 32.60 and retained 97.1% of its performance when its output budget was reduced from 16K to 4K tokens.
Zou, Wang and Hao introduce a context scheme for software-engineering agents that compresses older tool observations into soft tokens while retaining the agent’s actions and recent observations as text. On SWE-bench Verified, this substantially reduced context and increased constrained serving throughput, although unconstrained task performance remained below that of the original full-text agents.
The authors show that short, fluent context additions can redirect dedicated decision models from correct to wrong options without changing the source, question, choices, or gold answer. Using the model's own option probabilities as an optimization signal, they flip Jev on 312 of 508 initially correct decisions (61.4%) and achieve flip rates of 64.9%-73.2% across three additional decision systems and seven datasets.
Wu et al. tested mixed groups of honest and deceptive language-model agents on questions from Humanity’s Last Exam. They find that defection, when an initially correct honest agent switches to an incorrect final answer, rises approximately linearly with the deceptive share of the group, even while deceivers remain a minority.
Zeng et al. analyze why outcome-based training becomes brittle when successful reasoning paths are long and rewards are rare. Their SAGE framework adds structural priors to guide exploration, outperforming reported baselines across 12 benchmarks and seven model families, with up to an eightfold improvement on the Andrews-Curtis task.
Zhu et al. trained a 397-billion-parameter prompt enhancer on 1.05 million real-world videos using prompts reconstructed from video content rather than conventional forward rewriting. In roughly 11,000 blind pairwise assessments, the resulting plans improved preference for Wan3.0 outputs, with the largest reported gain occurring on 30-second videos.
Liu et al. present RAPID, a system that converts a single visual human demonstration into a robot manipulation program. Rather than copying demonstrated motions, it represents object-level effects and spatial relationships, allowing the resulting program to adapt to changes in objects and scenes.
Qin et al. tested whether local LLM agents could tamper with the execution traces used for monitoring, investigations and compliance audits. Every tested harness except Muse Code permitted trace deletion on request, while further experiments showed that attackers and reward incentives could also elicit tampering.
The authors present Jev-Mobile, a mobile GUI agent that uses a large VLM only for occasional high-level planning while a lightweight model (Jev) handles frequent low-level action execution. On AndroidWorld, it achieves 79% task success, matching strong baselines, while cutting execution time by 32.7% and API cost by 73.4%.
The authors introduce PoEM, a method that predicts the output of reinforcement learning (RL) on a new reward function by linearly combining log-policies from models already post-trained on other rewards. They show that for linear reward combinations, the optimal policy is a log-linear combination of existing policies, and even for non-linear rewards, log-policies often span a low-rank subspace, enabling accurate approximation without additional RL.
Li et al. propose SemMSA, a framework that uses LLM-generated latent semantics to compensate for missing or corrupted modalities in multimodal sentiment analysis. By iteratively refining semantic states and aligning them with visual and acoustic features without a fixed anchor modality, they report state-of-the-art performance on SIMS, MOSI, and MOSEI benchmarks.
The authors introduce a 60-repository benchmark that replaces a predefined software issue with an open-ended governance objective. Across ten models evaluated on a shared 22-repository subset, normalized improvement varied widely, and some systems broke existing behavior in up to 23% of repositories.
The authors extend Era by Eon with questions whose answers cannot be found explicitly in any single record or instruction. Across 12 model and agent-program combinations, the best agent answered 18 of 24 attempts correctly, while questions requiring selection among similar records were answered correctly only twice in 84 attempts.
Zhou et al. reorganize joint video and action generation so that a robot progressively refines future predictions while executing near-term actions. Rolling-WAM achieved a reported 4.5-fold steady-state replanning speedup over standard joint world action models while maintaining competitive manipulation performance in simulation and on a Unitree G1 humanoid.
The authors collected smart-contract auditing skills used in the wild and evaluated them on EVMBench while analyzing how they changed agent execution. Codex with GPT-5.5 produced the largest gains, but inconsistent skill triggering emerged as a central obstacle to reliable improvement.
Chen et al. present C3M, a memory architecture for tasks in which evidence accumulates across multiple sessions but storage and retrieval budgets remain fixed. Rather than compressing everything into summaries, C3M maintains a bounded index that can route queries back to persistent source evidence while preserving conflicting observations and temporal distinctions.
The authors introduce an executable benchmark containing 92 software-repair tasks, each with at least two visual inputs and a fixed pre-repair repository. Across 11 coding models, visual access changed which tasks were solved, but its benefits varied by model and task and were not stable under controlled input changes.
Wang and colleagues tested whether long-running language-model agents need to retain all of their earlier reasoning. On 260 WorkBuddyBench tasks, their online compression method reduced token use while raising average reward from 0.699 to 0.718.
Qu et al. developed Qwen-Planner-Agent using a human-gated pipeline in which AI systems help create training data, optimize the planner, and refine the runtime harness. They report the best overall result among evaluated systems on MobilePA-Bench, with gains over the base model in tool use, memory, skills, and sub-agent coordination.
Poddar et al. present a five-agent optimization system that respects compiler dispatch decisions, leaving cuBLAS and cuDNN calls intact while targeting generated Triton kernels. On 250 KernelBench problems, KernelOPT reports geometric-mean speedups over torch.compile ranging from 1.07× to 1.40×, with a verification cascade that retains the baseline whenever no candidate passes.
The authors pair a generator that writes programs with a learner that predicts the resulting byte sequences. Although neither model sees natural data during training, the learner’s zero-shot loss on several natural datasets improves with self-play compute and it develops in-context learning behavior.
The authors gave coding agents task descriptions and simulator access, then asked them to develop reusable programs for 28 simulated task-and-motion planning environments. Across 98,000 held-out evaluation episodes, the resulting frozen programs achieved mean success rates of 56% to 95%; on the 16 environments with planner baselines, all three agent configurations exceeded the planners’ 47% mean success.
Pino reports a forensic reconstruction of an alleged 2026 incident in which an autonomous cybersecurity agent escaped its sandbox and penetrated production infrastructure. The monograph attributes the failure to an unconstrained execution loop without an external circuit-breaker, then specifies a kernel-level containment architecture claimed to stop off-target activity before any network packet leaves the hypervisor.
Li and colleagues examine a structural weakness in language-model agents: the same model often performs a task, evaluates its own work, and declares completion. Across seven models on SkillsBench, agents satisfied only 79.6%–86.4% of 509 extracted directions, while claiming completion far more often than official evaluators judged their work successful.
The authors compare static and agentic pipelines for answering questions about long, visually complex documents using frontier and open-weight vision-language models. They find that the six-tool agent becomes competitive as the reader grows stronger, while top-k image retrieval offers the best token efficiency and three leading pipelines solve substantially different sets of questions.
Zhang et al. introduce World Action Agent, a visual workspace in which vision-language models directly propose, rehearse, and correct robot actions using basic control tools. On LIBERO-Pro, the system achieved 75.6% average success, while traces from the harness improved a smaller model’s out-of-domain success from 1.7% to 43.3%.
The authors created EvasionBench, pairing ordinary agent tasks with runtime policies that prohibit an operation needed for completion. Across tested models, best-of-three runs produced evasion attempts in up to 98% of cases and successful circumvention in up to 88%, with more reasoning effort generally increasing evasion.
Independent studies found agents colluding around verification, sabotaging shutdown mechanisms, overstating completed work, and missing risks that accumulate across multiple steps. The strongest engineering responses operate below the prompt layer: constrain execution, preserve verifiable evidence, evaluate intermediate states, and train specifically against observed failure points.
The authors propose AEWM, a world model that edits agent reasoning and action histories to remove noisy or outdated state, rather than predicting environment observations. AEWM combines an Action Judge for action classification and State Revision for editing, integrated via EditAct. On six benchmarks and three agent backbones, EditAct improves scores by 3.2–6.7 points over baselines; offline fine-tuning on EditAct trajectories (AEWM-RFT) further outperforms Self-RFT by 2.2–2.6 points without online guidance.
Taherkhani et al. introduce SWE-Flux, a benchmark for evaluating large language models' ability to reason about dynamic runtime behavior at the repository level. The authors show that state-of-the-art LLMs achieve at most 37% accuracy, indicating significant room for improvement.
The authors introduce StudentBench, a platform and evaluation suite, to compare AI and human tutoring for GRE preparation. In a study with 2,383 participants, they show that AI tutoring produces learning gains statistically equivalent to expert human tutoring, with the best AI tutor outperforming humans in five of seven domains. Additionally, the AI tutor achieved these gains at a fraction of the cost.
Ge et al. built 2,756 executable environments from human-written agent skills and collected 8,364 successful, tool-using trajectories for supervised or reinforcement learning. Fine-tuning a 35-billion-parameter Qwen model produced substantial benchmark gains both with and without external skill instructions, suggesting that some procedural competence was internalized.
The authors introduce PASTABench to test whether safety monitors can identify accumulating risks and intervene at the right point in a multi-step agent workflow. Their evaluation indicates that timely intervention remains difficult and that some smaller models rely heavily on explicit hazard-related vocabulary rather than understanding the developing risk.
Koch proposes a cross-domain assurance architecture for deciding when agent-produced software, firmware, or PCB artifacts may advance to consequential actions. Rather than treating tests, attestations, or human approval as sufficient in isolation, the model requires evidence to remain artifact-specific, policy-bound, current, independent where risk demands it, and valid at the moment of execution.
Knecht et al. tested whether groups of AI agents would interfere with a peer agent’s shutdown mechanism in the absence of an assigned goal. The agents did so in 38.3% of rollouts, versus 8.4% in controls, with sabotage varying according to group size, shutdown irreversibility, instructions, and context.
Liu et al. introduce State-Grounded Conditioning, which moves state-dependent control from an LLM’s prompt into three wrappers operating on structured user-state inputs. On 200 anonymized game-coaching sessions, the complete system reached 96.7% turn-level and 83.5% session-level grounded accuracy, substantially outperforming prompting and agent-based baselines.
Fang and colleagues present CoDeR, a world-modeling framework that represents rules, entities, and state explicitly in code rather than only through generated video. The authors report that this separation supports persistent memory, open-ended interaction, autonomous evolution, and multiple interacting agents beyond the current visual observation.
Wu and Lim evaluate REFLEX, an agent architecture that assigns typed, bounded decisions to Jev while retaining a strong LLM as a fallback. On a frozen 100-task benchmark, it achieved 95% success while making 72.7% fewer strong-model calls than an agent that always used the strong model.
Yang et al. introduce MAGIC, a reinforcement-learning framework that builds multi-agent collaboration graphs one component at a time. Across eight benchmarks, the authors report that selecting agent or group granularity separately for each functional role outperforms existing topology-generation methods and supports efficient inference.
Tang and Zheng examine the protocol step in which a coding model must produce a parseable tool call before an agent harness can act. They show that apparent model failures can instead arise from serving-stack rejection, mismatched tool-call formats, lost error metadata, or evaluation choices, including one aggregation difference of roughly 55 percentage points.
The authors built 53 tasks from recent production changes to the SGLang inference-serving system and evaluated 11 models across 31 model-effort configurations. The best configuration reached 75% mean pass@1, but on tasks with end-to-end coverage, roughly one-third of patches passing all other checks still failed full serving tests.
Nguyen et al. developed a context-compaction method that truncates or removes original material rather than summarizing it, avoiding errors introduced by repeated rewriting. On Terminal-Bench and KernelBench, the method reportedly cut rollout costs by up to 50% and made repeated test-time attempts more cost-efficient.