News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 15 of 54 · The Research front page
Han et al. study whether security checks can intervene after a reasoning model drafts code but before it delivers that code to the user. Across 173 benchmark tasks, their SecReason system substantially increased the rate of secure, functionally correct outputs on cases where unassisted models produced vulnerabilities, while using fewer tokens than post-generation repair.
Ling and colleagues tested whether replacing agents’ action-selection components with Jev, a model designed for fast choices among defined options, improves end-to-end efficiency. Jev lowered decision cost and latency in all five agents, but those local gains did not reliably translate into faster or more economical successful outcomes. Mobile agents were particularly vulnerable to failed recovery and declaring completion too early.
Wang et al. propose Jev-LiteMAS, a multi-agent framework that allocates routine coordination to a low-cost decision model (Jev) and reserves generative models for expert execution and corrective feedback. Across four benchmarks and two backbones, the method reduces generative-LLM tokens by 12.7–54.6% with accuracy within 1.34 points of the baseline, and lowers end-to-end latency in most configurations.
Wang et al. introduce SCCGBench, a benchmark of 3,135 real-world RapidAPI call functions mined from GitHub, designed to test whether LLM-generated code actually invokes the correct service endpoint. They evaluate six code LLMs and find that models often generate plausible-looking but non-executable code, with the best model achieving only 3.76% Mock success (hitting the correct host, endpoint, and HTTP method) without documentation. Providing gold API documentation boosts Mock to roughly 90%, highlighting that API grounding, not code generation, is the primary bottleneck.
The authors show that the same indirect prompt injection succeeds at dramatically different rates depending on the presentation channel, revealing a 'framing gap'. They propose TRACER, a channel-aware adversarial training method that significantly reduces attack success across channels on GSM8K with Qwen2.5 models.
Lai and colleagues introduce a pipeline for generating computer-use tasks in which agent messages conflict with evidence visible in the operating environment. Across 600 multi-round tasks and 15 models, stronger systems were more likely to inspect the environment, but detecting a conflict did not ensure that they corrected it or completed required handoffs.
Du and colleagues compared 21 models on 8,742 medical examples using a system-one protocol, meaning models answered directly without producing reasoning traces. Jev matched leading models on primarily text-based tasks and made unusually consistent decisions across formatting changes, but this consistency did not translate into reliable abstention or strong numerical performance.
Rajput et al. pre-registered and ran 299,463 API calls on Jev-1.13.0, a hosted decision model, to test whether its probabilities respect independence of irrelevant alternatives (IIA). They found systematic, directional shifts in log-odds between the correct label and its strongest rival whenever any option was added, even when that option carried less than 0.1% of the probability mass. The violations are substantial and structured, invalidating the assumption that probabilities are fixed properties of options.
Liu et al. compared three plan-based DataFrame code generation pipelines on 30 tasks derived from TPC-DS benchmark. They found that test-pass counts rose from 22/30 to 30/30 during revision yet no solution matched the hidden reference output, evidence of what they call interpretive lock-in: a mistaken interpretation persisting across plan, implementation, and tests.
The authors propose Large Decision Models (LDMs), which repurpose large language models to return explicit probabilities over labels in a single prefill forward pass. They combine diverse synthetic supervision with distillation of reasoning teacher outputs, and report consistent accuracy gains over a matched baseline while cutting median latency by a factor of 182 on Gemma-4-31B.
Zhou et al. present the first release of OpenGCUA, an open system for agents that choose among screenshots, GUI controls, shell commands and code when completing computer tasks. On an offline OSWorld-v2 evaluation, their full configuration improved all eight tested models, raising the mean partial score by 19.6 points and cutting reported steps per task by 53.0% relative to batched GUI-only execution.
Fang et al. add a dedicated decision head to an existing large audio language model, allowing it to answer bounded questions without generating text token by token. Across 18 evaluation configurations, the model handled classification, speaker verification and quality scoring while running approximately 20 times faster than its generative backbone in the tested inference setting.
Li and colleagues tested Jev, five general-purpose LLM judges and two local safety classifiers against identical adversarial payloads. Jev was comparatively robust when checking explicit rules, but on phishing, a task requiring an overall judgment, it detected fewer malicious emails and was susceptible to forged verdicts embedded in the input.
Seip, Markovinović and Hein introduce SA-CUA, a dual-LLM architecture for agents that operate graphical interfaces through screenshots, mouse clicks and keyboard input. In their experiments, deterministic masking based on accessibility metadata prevented untrusted free text from reaching the privileged model, while an independent verification channel preserved more task utility than the earlier CaMeL-CUA design and yielded zero successful attacks on the evaluated tasks.
Song et al. present LEV, a decision model that can vary its computational depth without generating intermediate reasoning text or loading a second model. Across three looped backbones, additional passes through the same layers improved accuracy primarily on difficult questions, while an adaptive stopping mechanism reduced computation by 15% to 24% with less than one percentage point of accuracy loss.
Duan, Lin and colleagues evaluated 25 configurations of general decision models and generative language models, then tested how their local predictions behaved inside longer workflows. The authors find that decision models compete well when the prompt contains the necessary evidence, but struggle with specialist knowledge, calibrated probabilities and sequences in which small errors compound.
The authors propose JEV-as-a-Judge, a minimal API that outputs only label probabilities. Evaluated on over 5,000 judgments, it trails GPT-6 overall but stays within three points on tasks where the verdict is explicit. When its confidence signals escalation, a cascade that mixes JEV and GPT-6 matches GPT-6's accuracy while using only 41% of its fee.
Luo and colleagues study a security failure they call Jev Program Decision Hijacking, in which untrusted context manipulates a Jev model into returning a structurally valid but semantically misleading judgment. Their automated framework, JevProbe, found program-level decision flips in 1,853 of 2,273 test cases, with nearly all of those flips linked to potential downstream security consequences.
Yinheng Li and Justin Wagle present LLM2Jev, a framework that converts any causal LLM into a Jev-style decision model by scoring bracketed numeric option IDs from next-token probabilities. On frozen Qwen3.5-4B, this training-free recipe matches or exceeds community Jev-style models on the same backbone and handles multimodal decisions. Fine-tuning improves weaker models and many-option routing, but yields diminishing returns for strong backbones.
Geuter et al. show that four open instruction-tuned language models can perform System One decision tasks, choosing among options in a single forward pass, with no additional training. After a short supervised LoRA fine-tuning recipe, Gemma 4 31B exceeds Jev on both OpenDecisions and JevBench, setting a new state-of-the-art.
The authors introduce P3-WAM, a unified world action model that trains 3D interaction reconstruction, action-conditioned world prediction, and joint action-future planning in one architecture. Unified training improves all three capabilities over single-mode training, and the model outperforms specialized baselines on hand reconstruction and action-conditioned video prediction across four datasets while producing goal-directed planning videos with stronger instruction adherence.
The authors introduce SUPRA-Tex, a framework for single-image 3D generation that separates material estimation, surface alignment, and geometry-conditioned completion. It projects reference-view PBR maps into the target latent space and uses an interleaved flow transformer to complete the full surface. Trained only on public data, it matches or exceeds commercial systems on in-the-wild images.
The authors propose GeoState, a view-decoupled geometry representation that disentangles scene content from camera viewpoint when using pretrained geometry foundation model (GFM) features as predictive states. By aggregating multi-view GFM features into a shared scene state and querying it only at decoding time, the model allows a temporal predictor to focus on genuine scene evolution. GeoState outperforms existing 3D world models on Waymo in depth forecasting, supports geometrically consistent cross-view queries, and generalizes zero-shot to KITTI and nuScenes.
The authors introduce KnowGPO, a reinforcement learning framework that assigns credit in deep search agents by representing each step's state as a subset of atomic knowledge conditions satisfied, rather than by surface observations. Trajectories with the same knowledge state are grouped for within-group comparison, and the incremental increase in satisfied conditions between steps serves as a dense process reward. On three benchmarks (Xbench, BrowseComp-ZH, GAIA), KnowGPO improves over outcome-only GRPO by 2.0 to 5.9% and exceeds recent process-supervision methods by 3.8 to 4.2 percentage points, with an 8B model matching or surpassing most 32B open-source systems.
Chavan et al. introduce Subspace Tuning (SST), a method that restricts each weight update to the bottom singular subspace of the weight matrix, preventing changes along the principal directions. Across SFT and self-distillation, SST preserves prior capabilities (e.g., reasoning, tool use) while matching or exceeding the target-task accuracy of unconstrained fine-tuning, with 2–4× less KL divergence from the base model.
The authors propose Routing Policy Optimization (RPO), a method that introduces stochastic exploration into diffusion and flow generators by randomly sampling attention routes during training while preserving the original deterministic sampler at inference. RPO uses a Gaussian velocity likelihood to score perturbations and a controlled two-rollout estimator to compute advantages, achieving better reward-aligned generation quality on text-to-image and text-to-video tasks while reducing training cost by up to 4.2x compared to Flow-GRPO.
Zhang and colleagues present a pipeline for training multimodal agents to investigate questions through multiple rounds of web search, visual retrieval and reasoning. Their 9-billion-parameter agent improved average accuracy by 15.3 percentage points over its agentic-workflow baseline across seven primary benchmarks, with joint reinforcement learning and on-policy distillation adding 2.3 points beyond supervised fine-tuning.
The authors propose SCOPED, a framework that trains separate capability-specific experts via reinforcement learning, extracts explicit skills from their reasoning trajectories, and uses those skills to condition the experts during on-policy distillation into a unified student vision-language model. Across eight multimodal benchmarks, SCOPED enables a compact 8B model to achieve the highest average score among evaluated models, surpassing both the backbone and mixed-capability RL baselines.
The authors organize specialized LLM agents into a market where each agent buys an intermediate task state, transforms it and auctions the result onward. Across synthetic and LLM-based tasks, they find that moderate competition can support both positive agent profits and high task success, while monopoly-like markets fail and perfect competition eliminates agents’ margins.
Fei and colleagues trained Hy-Memory-30B-A3B to decide what information an external memory system should store, update or discard. On fixed 20% benchmark subsamples, the model ranked first across four evaluation harnesses, posting a 55.49% macro average compared with 54.62% for Claude Opus 5.
Qian, Ren and Bai augment feedforward transformer models for multi-view 3D reconstruction by supplying optical flow and gravity direction as privileged information during training only. Evaluated on standard benchmarks, the models show consistent gains in accuracy and robustness, especially for dynamic objects, ambiguous structure and challenging camera motion. The optimal injection stage is architecture-dependent: middle-layer integration suits VGGT, while late-stage integration suits π3.
The authors propose EMO, a training-free framework that alternates question-specific context optimization (E-step) with confidence-constrained logits steering (M-step) while model weights stay frozen. Across eight models and six benchmarks, it improves zero-shot accuracy by 9.10 percentage points on average, and adds 1.53 points over context-only optimization with the same judge.
Zhang Xiaofeng and colleagues separate token learning from pixel rendering: the tokenizer is trained to reconstruct semantic features, while a detached decoder converts tokens into images without letting pixel losses reshape the representation. Across four pretrained vision encoders, this produced tokens that image generators could model more effectively, improving generation quality at substantially lower sampling compute than evaluated continuous alternatives.
DropKV addresses the KV-cache memory bottleneck in long-context LLMs by turning the NP-hard problem of minimizing output perturbation from cache eviction into independent per-token scores. The authors prove this decoupled proxy is within a constant factor of the optimal combinatorial solution under a cone condition they verify empirically, and implement it in fused Triton kernels. Across RULER and LongBench at 5% and 10% budgets, DropKV beats existing eviction baselines in accuracy and latency.
The authors introduce Intern-Decision, a family of multimodal decision models (0.8B, 2B, and 4B parameters) that reformulate structured decision making as next-token prediction over masked fields in a predefined JSON skeleton, enabling multi-field scoring in a single causal forward pass. The 4B model achieves a 90.02% unweighted average accuracy across seven evaluation suites, compared with 88.74% for the baseline Jev model, and attains a mean latency of 44.16 ms for a 289-token request with three decision fields on one RTX 4090.
The authors tested eight methods for substituting a sender's KV cache into a frozen reader LLM in place of its own cache. All substitution methods performed at the level of an uninformative cache baseline, failing to improve the reader's accuracy. In contrast, an augmentation approach—where the reader retains its own cache and attends over the sender's cache as additional context, with only its queries adapted—beat the reader's native-context baseline and worked when pooling multiple senders.
The authors introduce AgentPerturb, a framework that applies missing, redundant, and incorrect edits to user utterances, procedural specifications, carried-over context, and tool contracts in agentic tasks. Evaluating ten models on APEX-Agents and Toolathlon, they find consistent performance drops, with vulnerability organized more by perturbation target than type, and successful perturbed trajectories often requiring substantially more computation.
Muon optimizer achieves lower validation and test loss than AdamW when given the same number of training tokens on small language models and vision transformers. However, its per-step overhead is substantially larger, making it slower in wall-clock time for these small models. The authors characterize conditions where Muon wins or loses under data, time, and tuning budgets.
Floros et al. identify a flaw in how classifier-free guidance (CFG) is applied to text-and-image-to-image (TI2I) flow matching models. They find that CFG's unconditional branch approximates the average of all possible edits, which harms source preservation during the noisy early phase of generation. They propose a simple inference-time correction: substitute a no-edit velocity into the unconditional branch for early flow steps. This intervention recovers a richer alignment-preservation tradeoff without additional model evaluations and is validated on both pixel-space and latent TI2I flows.
The authors introduce Localized Context Distillation (LoCD) to improve how large language models learn from interaction experience. LoCD first localizes the spans of a student trajectory that require revision, then has a teacher regenerate only those spans from the student's own prefix, and finally trains the student on the revised segments. In extensive experiments on math reasoning, code generation, and interactive agent tasks, LoCD consistently outperforms strong baselines while requiring smaller and sparser parameter updates.
The authors build AgentDecisionBench, JevAgentHarness, and AgentDecision-4B to test when a fast decision model can take over routine choices inside an LLM agent. They find that a 4B model trained on the agent's own states solves 91.5% of out-of-distribution tasks versus 82.0% for the stronger Jev model, and the harness cuts agent cost by 81.1-89.0% and latency by up to 2.5x.
The authors propose TRACER, a retrieval front-end that prunes tool registries to a small candidate set by exploiting the sequential nature of tool calls in multi-turn agent tasks. On a 128-tool retail-and-telecom benchmark, TRACER improved GPT-5.4's tool-selection accuracy by 4.7 percentage points while reducing input tokens per step by 80%, presenting only 5 tools at a time.
The authors propose a framework for token-level teacher selection during on-policy distillation (OPD), where different teachers can supervise different tokens of a student rollout. They prove conditions for teacher-student compatibility and introduce representation-based routing, which outperforms confidence-based methods in producing diverse, high-quality student responses. Experiments across math, code, kernel generation, and visual reasoning show consistent improvements of up to 10 percentage points in pass@k over vanilla OPD and individual teacher experts.
The authors investigate whether latent chain-of-thought models, which use continuous hidden states instead of explicit reasoning steps, actually perform reasoning. They find that on logical tasks, the latent chain often carries a pre-existing answer, and deleting the chain leaves predictions unchanged. A new taxonomy is introduced to clarify what a chain represents versus how it is used.
The authors propose CleanRM, a reward modeling method that estimates sample-wise noise ratios from a small set of clean preference labels using Wasserstein optimal transport. They prove the resulting training objective is unbiased with respect to the clean risk, and show that it outperforms standard denoising baselines and improves downstream RLHF alignment.
Dai et al. address a central obstacle in self-generated training for tasks without objectively checkable answers: models can exploit incomplete grading rubrics rather than produce genuinely useful work. Their Recursive Reward Improvement method adversarially tests and repairs each rubric before training, producing substantial gains on medical benchmarks and some transfer to unseen professional domains.
The authors introduce a training method that teaches an 8B language model to generate complete GPU optimization scripts and improve existing ones. Trained on search-derived examples from TVM MetaSchedule, the model reaches 91.27% correctness for construction and 92.01% for refinement, and produces faster kernels than six frontier LLMs with hundreds of billions to trillions of parameters.
Cicek and colleagues present TEAL, an active-learning method for training semantic segmentation models with extremely sparse pixel annotations. Across four datasets, it outperformed prior approaches, with the largest gain reaching 24.36 mean intersection-over-union points, and recovered up to 97% of fully supervised performance using 10 labeled pixels per image.