Translate
Live intelligence

Latest AI updates

573 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research Editor selected

LoCo Maps Where Concepts Emerge in Visual Autoregressive Models

The paper introduces LoCo, a model-agnostic framework that localizes visual concepts across layers, generation scales and positions in next-scale autoregressive models. It also presents LoCoBench, a dataset spanning 10 categories, and reports evaluations on image and video models including Infinity, HunyuanImage-3.0 and InfinityStar. The authors report causal links between localized positions and concept emergence, with applications to concept erasure, personalization and adversarial concept injection.

Why it matters: The work offers a concrete approach to understanding and selectively modifying visual generative models, although its performance claims require scrutiny beyond the abstract.

arxiv6d agoScore 83
LoCoLoCoBenchVisual autoregressive models
Research

Agentic Process Evidence: Signed Attestations for AI Agent Sessions in the SDLC

论文指出 AI 代理现在承担创建 PR、选择依赖、重构代码库、解释测试失败以及执行基础设施变更等任务，但会话一旦结束，它们的推理、决策和工具调用记录就会消失；而 git log 只显示改了什么，不显示哪个模型、调用了哪些工具、谁对结果负责。作者提出 Agentic Process Evidence (APE)，这是一种签名、防篡改、机器可读的代理过程证明，把代理会话与其结果绑定。APE 支持 in-toto 风格证明框架，并支持 Policy-as-Code 控制，只在需要时引入人工介入。论文还指出三个程序性问题：供应链向量、策略护栏和任务对齐。

arxiv6d agoScore 68
AI agent governancesoftware supply chain securitySDLC automation
Research

DS-Upcycling Expands Dense LLMs into Domain-Specialized Experts for Multi-Domain Adaptation

The paper proposes Domain-Specialized Upcycling (DS-Upcycling), which expands a dense Transformer layer with additional experts and a router rather than attaching per-domain LoRA adapters. A new domain-aware supervision signal trains the router to select the right expert for each input domain, and fine-tuning on a new domain updates only its matching expert and the router while freezing everything else. Across adapted domains the method reports the highest average performance among baselines while preserving general capability. With only 256 examples per domain for router supervision, routing accuracy after adaptation reaches 90-98%, versus 43-70% for routers trained without the proposed approach.

arxiv6d agoScore 74
Model UpcyclingMulti-Domain AdaptationMixture-of-Experts
Research

RACE: Reward-Adaptive Co-Evolution of Policy and Rubric for Open-Ended RL

This arXiv paper introduces RACE, a framework that co-evolves a reinforcement-learning policy and the rubric used to score open-ended responses. At the group level, per-rollout-group signal routing classifies groups as informative, diversity-collapsed, or rubric-failed, and only informative groups update the policy while failures are cached. At the phase level, an event-triggered switch moves into rubric optimization when rewards saturate and returns when rubric-reward gains plateau, with online proportional control adapting both thresholds. The authors report the highest average performance under a unified protocol, beating the strongest fixed-alternation baseline on all eight tasks and improving rubric discrimination.

arxiv6d agoScore 68
RACErubric-based rewardpolicy optimization
Research

BGP: Branch-aware Graph Planning Targets "Path Mirage" in Offline Goal-Conditioned RL

A new arXiv paper addresses offline goal-conditioned reinforcement learning in stochastic long-horizon settings, where compounding value-estimation errors make reliable goal reaching difficult. The authors identify a failure mode they call "path mirage," in which agents overcommit to spuriously successful trajectories that stochasticity induces in offline datasets. Their method, Branch-aware Graph Planning (BGP), builds graphs by treating high-variability points as nodes and learns goal-conditioned edges with segmented learning to avoid unreliable branches and cut stochasticity along paths. Experiments on diverse OGBench tasks reportedly outperform prior state-of-the-art offline goal-conditioned RL methods, and the implementation is publicly linked.

arxiv6d agoScore 71
offline reinforcement learninggoal-conditioned RLgraph-based planning
Research Editor selected

DynaRev Benchmarks Web Agents on Changing User Goals

DynaRev introduces 181 tasks that evaluate how web agents respond when users revise requirements during execution, measuring correctness, responsiveness and efficiency. Evaluations across five models show that agents often fulfill only part of the requirements, with results varying by task source, revision type and model. Further analysis finds gains from rewriting instructions while preserving browser state, but losses from browser resets; code and data are available.

Why it matters: The benchmark tests agents’ ability to handle evolving user requirements, exposing limitations that fixed-goal evaluations may miss.

arxiv6d agoScore 85
DynaRevWeb agentsAgent evaluation
Research

FARE: Factor-Aware Retrieval with Error-Guided Fusion for Retail Sales Forecasting

FARE is a retrieval-augmentation method for short-term retail demand forecasting, where target weeks often combine promotion and calendar factors that have few exact historical precedents. It ranks historical weeks separately for each known target-week factor, merges the top matches into a factor reference under a fixed budget, and blends that with a shape reference drawn from similar sales histories. Two gating mechanisms, driven by shrunk errors from completed development weeks and by current disagreement among candidates, set the mix and control how strongly it corrects a backbone forecast; both gates stay fixed at test time. Across Favorita, M5, Rossmann and Rohlik, FARE cut WAPE of five backbones by 11.46%, 4.06%, 14.63% and 10.16% on average, and also beat supervised residual correction and retrieval augmentation with Chronos-2 frozen. Code is released publicly.

arxiv6d agoScore 76
Retail demand forecastingRetrieval-augmented predictionTime-series foundation models
Research

Negotiation-Aware Hybrid LLM Framework Tackles Conflicting Cloud Intents

The paper proposes a framework that refines natural-language cloud intents, dynamically detects conflicts, and structures the result into a provider-independent TOSCA template. Candidate service offers are then retrieved through deterministic matching grounded in provider catalogs. Conflicts found during interpretation and requirements that prove unsatisfiable during discovery are resolved through iterative natural-language negotiation with the user. On 70 diverse intents the framework improves recall and F1 for both conflict detection and service discovery over single-shot and enhanced-prompting baselines, and eliminates hallucinated offers.

arxiv6d agoScore 72
LLM agentscloud service compositionintent understanding
Research

EventSeeColor Uses Diffusion to Reconstruct Color from Monochrome Camera Events

EventSeeColor combines event-reconstructed grayscale images with aligned event data to guide a diffusion-based colorizer. The method uses semantic context, region-balanced chroma supervision and local RGB refinement to improve reconstructed appearance. The authors report leading structural similarity, perceptual quality and chromatic-region color accuracy among the external methods evaluated across two real event datasets.

arxiv6d agoScore 79
EventSeeColorEvent-based visionDiffusion models
Research Editor selected

SPLICE Locates and Closes Information Gaps in Multi-Agent LLM Workflows

The paper addresses a structural weakness in fixed multi-agent workflows, where evidence gathered by one agent may never reach the agent that needs it, producing an information gap and an incorrect answer. SPLICE runs the original workflow, locates a likely gap, inserts one directed edge from a source to a receiver, and re-executes the receiver and affected downstream nodes to produce a candidate answer. A selector trained on process-level features, such as how far the receiver moved and whether the change propagated to the final output, decides whether to adopt the candidate or keep the original answer. Across five benchmarks, SPLICE improves results on Web tool-use tasks by 8.0% and the mean score on four reasoning and planning benchmarks by 5.7% at 1.25 to 1.44 times the skeleton's cost, and it transfers zero-shot to four unseen backbones.

Why it matters: It offers a concrete, cost-characterized method for diagnosing and repairing missing evidence in multi-agent pipelines, with measured gains and zero-shot transfer evidence.

arxiv6d agoScore 82
Multi-agent LLM orchestrationAgent workflow optimizationInformation flow in agent graphs
Research

Theoretical Analysis of Axial Attention in Time Series Foundation Models

The authors extend the Single Location Regression framework from Marion et al. [2025] to 2D structured data in order to study sparse-information settings where a target depends on only a small number of tokens. They formulate an analytically tractable axial attention-type predictor and analyze its statistical properties and training dynamics. Given oracle parameters in a specific asymptotic regime, the predictor is proven to achieve Bayes optimality, whereas a class of semi-linear simplifications provably fails, highlighting the role of the inner non-linearity. They also prove that Projected Gradient Descent converges globally to the optimal oracle parameters under a specific temperature scaling, and numerical experiments emphasize the importance of inverse temperature scheduling for convergence in practice.

arxiv6d agoScore 68
time series foundation modelsaxial attentionstatistical learning theory
Research

Bridge-LP: A Theoretically Grounded Framework for Link Prediction on Text-Attributed Graphs

Researchers present Bridge-LP, a cross-modal framework that unifies graph and language representations for link prediction on text-attributed graphs. The design combines textual-proximity jumping connections, compressed embeddings that encode neighborhood structure, and a cache embedding training strategy to reduce computational complexity. In experiments against 40 existing models across 9 datasets, Bridge-LP reports consistent state-of-the-art gains, including up to 43.06% improvement in Mean Reciprocal Rank.

arxiv6d agoScore 78
Graph Neural NetworksLanguage ModelsLink Prediction
Research

EpiAgent: Epistemic Self-Play for LLM Agents

A new arXiv paper introduces EpiAgent, a framework that gives LLM-based agents an explicit, learnable epistemic layer holding beliefs about other agents' private information. The agent's policy is split into two coupled stages, Epistemic Update and Action Planning, whose dependency is formulated as a bi-level optimization problem. The authors propose Epistemic Self-Play, an algorithm that alternately optimizes both stages via self-play using an intrinsic prediction reward measuring how well inferred beliefs predict later observations. The abstract reports that EpiAgent outperforms prompting-based and RL-based baselines in mixed and competitive scenarios, with analyses pointing to gains in belief inference accuracy and in action planning from learned beliefs.

arxiv6d agoScore 68
LLM agentsepistemic reasoningself-play
Research Editor selected

EnvOS Audits Computer-Use Agents on Mid-Task Changes

EnvOS is a research framework that checks whether mid-task environment changes execute correctly, tasks remain solvable, and graders distinguish correct from incorrect behavior, then independently rechecks verdicts against archived evidence. In the authors’ clinic-booking tests, Claude Opus 5 succeeded on 0 of 16 tasks without extra information, 2 of 16 with a reminder to recheck, and 15 of 16 when shown the updated value. Results varied by model and interface; on a separate externally generated set of corrupted episodes, the auditing pipeline caught 25 of 39 cases versus 4 for simple log checks.

Why it matters: The work offers a concrete method for separating agent failures from environment and grading failures, while showing how interface design can affect handling of stale information.

arxiv6d agoScore 86
EnvOSComputer-use agentsAgent evaluation
Research

Reaction Precedents Extend Retrosynthesis Beyond Training Coverage

Researchers used 2.71 million reactions to test whether retrieving reaction precedents can compensate for gaps in task-specific training coverage. With model parameters fixed, expanding only the retrieval corpus raised mean Top-1 accuracy on excluded transformations from 13.0% to 58.0% on a historical test set and from 4.9% to 23.1% on a temporally held-out set. A model trained on pre-2017 records matched or exceeded a non-retrieval model trained on pre-2025 records after retrieval expansion. Without task-specific fine-tuning, precedents improved all 14 representative general-purpose models tested, and the strongest model with precedents surpassed all eight representative specialized models.

arxiv6d agoScore 78
retrosynthesisretrieval-augmented generationAI for science
Research

TACE: Two-Stage Advantage and Credit Estimation for Multimodal RLVR

The paper analyzes advantage estimation in GRPO and factors it into sample-level baselining and token-level credit assignment, noting that existing estimators typically handle only one stage. TACE replaces linear standardization with a one-dimensional optimal-transport map to a standard normal for baselining, then rank-calibrates a stop-gradient evidence contrast and renormalizes clipped token weights to mean one. The authors prove inter-task moment equity, a deterministic tail bound, and mass-conservation, boundedness and sign-anchoring properties. On a multimodal reasoning suite TACE is strongest on most benchmarks but does not dominate every setting.

arxiv6d agoScore 74
GRPOreinforcement learning with verifiable rewardsmultimodal reasoning
Research

TaskFrame Connects Task-Level Planning with Frame-Level Control for Game Agents

TaskFrame is a hierarchical planner-controller framework that pairs a vision-language task planner with a causal Transformer controller that emits keyboard and mouse actions at every game step. The planner decomposes a high-level goal into an ordered sequence of atomic instructions and is optimized with supervised learning plus dependency-aware GRPO, while hidden states of goal tokens provide cached goal embeddings reused across control steps. Training runs in three stages: planner training, controller pretraining, and joint alignment. In Minecraft, the system is competitive with general-purpose vision-language models on instruction decomposition and reaches a higher closed-loop combat success rate than the evaluated baselines.

arxiv6d agoScore 74
TaskFramegame agentsvision-language planning
Research

Neighborhood Elites: Quality-Diversity Search for Uncertainty Quantification in Question Answering

The paper recasts perturbation-based uncertainty quantification for question answering as a quality-diversity search problem. Its proposed Neighborhood Elites (NE) constructor spreads answer-preserving rewrites across two controlled axes: the kind of edit made to the context, and how explicitly the context supports the answer. Across five extractive-QA datasets and five instruction-tuned LLMs, NE improves failure detection over three other neighborhood constructors in at least 24 of 25 dataset-model pairs, reaching 0.726 AUROC. It also outperforms semantic entropy computed over decoder samples, with the gap widening as the neighborhood grows.

arxiv6d agoScore 74
uncertainty quantificationquestion answeringquality-diversity search