Live intelligence

Latest AI updates

50 updates

Research

World-Model-Guided Navigation Distills Foresight into Fast Policy

Researchers propose a training-time approach for human-like navigation in multi-agent scenes, using world models as preference oracles to distill foresight into a fast policy. Two task-aligned world models score candidate actions by predicted consequences, and the policy is refined via group-relative preference distillation. At deployment, the world models are removed, enabling single forward-pass action selection. Evaluations across five scenarios show safer, more socially coherent, and more human-like navigation than baselines.

arxiv14h agoScore 74

world modelsrobot navigationreinforcement learning

Research

AttnSR: Attention Step Reward for Long-Horizon Agentic RL

AttnSR is a step-level reward augmentation method for long-horizon LLM agents that derives reward allocations directly from policy attention during training replay, eliminating the need for a separate reward model or additional rollouts. By tracing backward through the policy model's attention graph, it estimates the relevance of earlier steps to the final response and uses a running baseline to determine whether rewards reinforce or penalize selected steps. In experiments with Qwen2.5-7B-Instruct, AttnSR improved ALFWorld success from 91.2% to 97.1% and consistently outperformed PPO on ScienceWorld while reducing environment interactions by 29.9%.

arxiv14h agoScore 75

reinforcement learningLLM agentsreward shaping

Research

RM-VLM uses vision-language models to ground LLM-generated reward machines

Researchers propose RM-VLM, a reinforcement-learning approach that has a large language model generate a reward machine from a natural-language task description. A vision-language model then serves as the labelling function, grounding the machine's propositions in image-based observations during policy learning. A VLM-based value estimator assesses progress toward reward-machine transitions and produces a dense reward signal. The method is evaluated on non-Markovian grid-world navigation, robotic manipulation, and 3D object collection, where it reports higher task success rates than existing VLM-based reward methods.

arxiv14h agoScore 75

reward machinesvision-language modelsreinforcement learning

Research Editor selected

RLVR-Optimized Icon Generation for Alexa+

This paper describes a production pipeline for generating animated icons for Alexa+ devices from text descriptions, replacing a slow hand-drawn process. The authors focus on improving the critical image generation stage using reinforcement learning from verifiable rewards (RLVR) with objective verifiers that enforce design guidelines. The fine-tuned model increased the pass rate from 15.7% to 58.7% under an independent VLM judge, and both the model and verifier are deployed in production.

Why it matters: This is a concrete, recent deployment of RLVR in a production system, showing significant quality improvements and practical value for AI-driven content generation.

arxiv1d agoScore 82

reinforcement learningimage generationproduction deployment

Research Editor selected

Rubric Curriculum RL Improves RLVR in Non-Verifiable Domains

The paper introduces Rubric Curriculum RL (RcRL), a method that exploits the generation-verification gap in domains without ground-truth rewards by combining pairwise preferences over rollouts with a curriculum over rubric criteria. This creates a more discriminative and non-stationary training signal, avoiding the plateau seen in baselines. RcRL achieves a 70.5% human win rate on creative writing and improves HealthBench performance by 14.6% (full set) and 25.4% (consensus subset), outperforming all training baselines.

Why it matters: This presents a concrete, novel training method with measurable gains in open-ended domains, addressing a key limitation of RLVR.

arxiv1d agoScore 82

Reinforcement LearningReward ModelingLLM Alignment

Research Editor selected

Traffic-VL: Reinforced Vision-Language Model for End-to-End Traffic Signal Control

Traffic-VL is an 8-billion-parameter vision-language model that directly maps multi-view surveillance images and natural-language phase descriptions to traffic signal decisions, bypassing intermediate structured states. It is trained with offline reinforcement learning on photorealistic synthetic data (SynTra-V) and online stepwise trajectory optimization, matching structured-state baselines while outperforming image-based controllers and general-purpose VLMs. Evaluations include unseen topologies, degraded sensing, real footage, and a three-week deployment across ten intersections.

Why it matters: This paper presents a concrete, recent AI development with strong experimental validation, including real-world deployment, offering practical value for autonomous traffic management.

arxiv1d agoScore 85

Traffic signal controlVision-language modelsReinforcement learning

Research

GIST: Gibbs-Targeted Hierarchical Data Selection for Offline RLVR

This paper introduces GIST, a two-level data selection method for offline reinforcement learning with verifiable rewards (RLVR). It derives a factorization of the KL divergence between the current policy and the Gibbs-optimal target, identifying which responses and tokens are most informative. GIST selects correct responses on-policy and incorrect responses near the policy's mode, and supervises only the tokens where the mismatch lies. Experiments show accuracy gains across GSM8K, MATH-500, and MBPP for three models, with 66.6% fewer supervised tokens in a controlled comparison.

arxiv1d agoScore 78

Reinforcement LearningData SelectionToken Efficiency

Research Editor selected

Online Rubric Self-Distillation for Token-Level Credit Assignment in RL

The paper introduces ORSD, a method that generates natural-language process rubrics from on-policy trajectories to provide dense token-level supervision for reinforcement learning. A frozen verifier selects the best rubric bundle based on alignment with task success, and the likelihood gap with and without the rubric yields token-specific credit added to trajectory-level advantages. Experiments on ALFWorld and WebShop show faster policy improvement and higher final performance for Qwen3-1.7B and Qwen3-4B.

Why it matters: This is a novel, concrete AI research contribution with empirical results, addressing a key challenge in RL credit assignment.

arxiv1d agoScore 85

reinforcement learningtoken-level credit assignmentprocess rubrics

Research

CANDO: Conservative Action-Noise Directional Optimization for Residual Reinforcement Learning

This paper introduces CANDO, a residual reinforcement learning method that improves frozen generative policies by combining noise-space and action-space residual optimization. CANDO removes fixed latent feasible regions and uses conservative value estimation with gated execution to regulate adaptation. Experiments show CANDO achieves 76.1 NAUC and 91.4% final success rate, surpassing single-space baselines by 6.6 NAUC and 4.1%.

arxiv2d agoScore 78

Reinforcement LearningRobot ControlPolicy Adaptation

Research

ReSI-Math: Recoverable Capability Frontiers for Efficient Reasoning Improvement

ReSI-Math is a framework that identifies recoverable reasoning failures in LLMs—cases where direct decoding fails but trajectory-level search can elicit correct solutions—and converts them into efficient RLVR supervision. Using self-judge-guided evolutionary search, it selects frontier problems for sequential RLVR stages. On Qwen3-VL-8B, it improved MATH500 and WE-MATH averages from 78.71 to 80.16 after three stages, and frontier-guided selection outperformed random selection in matched single-problem RLVR.

arxiv2d agoScore 78

reinforcement learningLLM reasoningdata efficiency

Research

Density-Aware Reward Scaling for Diverse High-Reward Outcomes in RL

DARS is a lightweight reward-processing method that inversely scales rewards by local outcome density to promote reward-proportional sampling in continuous spaces. It requires no learned density model or auxiliary objective)Skip, and works with standard policy-gradient training. In multimodal navigation and text-to-image prompt adaptation, DARS-GRPO avoids mode collapse and improves diversity while maintaining semantic alignment.

arxiv2d agoScore 78

Reinforcement LearningReward DesignDiversity in AI

Research

FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation

FourSight addresses instability when applying NVFP4 W4A4 quantization to reinforcement learning post-training. It identifies that failures concentrate in outlier sequences and high-variance activation channels, then applies TIM-aware sequence filtering and sparse BF16 channel compensation. The method sustains near-BF16 accuracy and improves step-to-accuracy by 1.76-2.60x while staying within 2.3% of FP4 step time up to 120B-scale models.

arxiv2d agoScore 78

quantizationreinforcement learningmodel deployment

Research Editor selected

Artifact-to-Capability: Diagnosing Breaks in Capability Formation of Self-Evolving Agents

This paper introduces a formal definition of capability formation from artifacts in self-evolving agents, proposing five sequentially dependent properties and a three-scale diagnosis method. In freestyle Gomoku, Reflexion, Voyager, and DGM all show standard proxy signals but exhibit distinct capability-formation breaks, such as legalization gain, compositional override, and loss-only drift. The findings demonstrate that signal-based evaluation cannot reliably determine whether a capability has formed.

Why it matters: Provides a concrete diagnostic framework and counterexamples that challenge common evaluation practices in self-evolving AI agents, offering practical value for researchers and developers.

arxiv2d agoScore 85

self-evolving agentscapability evaluationagent diagnostics

Research Editor selected

WebCatalyst: High-Fidelity RL for Interactive HTML Generation

WebCatalyst is a reinforcement learning framework that improves interactive HTML generation from natural-language specs, using a game-level GUI assessor, Visual Anchor to prevent reward hacking, and Action-Trace Credit for token-level credit assignment. Trained on under 1K requests, it raises Cookie-Bench Mini Games from 68.8 to 75.3 and WebGen-Bench accuracy from 48.8 to 56.9 on Qwen3.6-35B-A3B, outperforming Claude Opus 4.7 in average rank. The work highlights feedback fidelity as key to advancing interactive HTML generation.

Why it matters: Provides a concrete, novel RL method with measurable benchmark gains and analysis, valuable for AI researchers and practitioners in code generation.

arxiv2d agoScore 82

WebCatalystinteractive HTML generationreinforcement learning

Research Editor selected

JHiMA-WebRL: Joint Hierarchical Multi-Agent RL for Long-Horizon Web Tasks

JHiMA-WebRL introduces a hierarchical multi-agent reinforcement learning framework for web agents, coupling a task decomposer with an actor and jointly training them. Using teacher-generated subtask supervision and an extension of GRPO, the framework achieves 37.2% success on WebArena-Lite-v2 with a 14B model, outperforming a 32B monolithic RL agent. Ablations confirm that joint updating of the decomposer is key.

Why it matters: This paper presents a concrete, recent AI development with clear performance gains and a novel joint training method, offering practical value for long-horizon web automation.

arxiv2d agoScore 85

Reinforcement LearningWeb NavigationMulti-Agent Systems

Research Editor selected

PEARL: A Durable Method for Suppressing Misaligned Behaviors in Language Models

This paper introduces PEARL, a GRPO-based reinforcement learning variant that suppresses undesired model behaviors more durably than standard methods. It tests on two deliberately misaligned model organisms with narrow and broad triggers, showing strong immediate suppression, preserved task accuracy, and durability under further fine-tuning. The method is most beneficial for broad misalignments that resist targeted SFT.

Why it matters: Provides a concrete, novel method with empirical results on durability, a key challenge in AI safety.

arxiv2d agoScore 85

AI alignmentbehavioral suppressionreinforcement learning

Research Editor selected

What Drives the Zero-to-One Jump? Diagnosing and Steering Hard-Prompt Learning in RLVR

The paper investigates the abrupt transition from unsolved to solved states in reinforcement learning with verifiable rewards (RLVR) for mathematical reasoning. The authors identify that negative rollouts preceding successful transitions retain higher early-stage generation entropy, motivating a new method called NAPO (Negative-Augmented Policy Optimization). Integrated with GRPO, NAPO injects informative gradients from negative examples and achieves significant improvements on AIME24 and AIME2025 benchmarks under fixed rollout budgets.

Why it matters: Provides a concrete diagnosis and steerable method for the zero-to-one jump phenomenon in RLVR, offering practical gains on hard math benchmarks.

arxiv3d agoScore 88

reinforcement learningmathematical reasoningpolicy optimization

Research Editor selected

Offline Approximate Value Iteration: A Reinforcement Learning Approach for Satellite Constellation Deployment

Researchers introduce a reinforcement learning-based approximate dynamic programming method to optimize Walker-type satellite constellation deployment under uncertainties like launch failures and satellite malfunctions. The approach uses offline approximate value iteration with linear function approximation and action pruning to reduce computational demands. Preliminary results show lower expected costs and deployment times compared to a rolling-horizon baseline, demonstrating a novel application of approximate dynamic programming in space logistics.

Why it matters: Presents a concrete, novel RL application to a real-world space logistics problem with measurable improvements over baselines.

arxiv3d agoScore 85

reinforcement learningspace logisticssatellite deployment