Tag
WeVisCut 是一个脚本到视频的智能体系统，通过训练一个 9B 紧凑策略来协调视频编辑工具、从故障中恢复，并在保持质量的同时大幅降低 token 消耗，在长视频生成上以约 33 倍更少的 token 达到与 Claude Opus 5 相当的性能。
该论文构建了一个最小化的策略学习模型，从数学和数值上分析了课程学习在奖励稠密与奖励稀疏情形下的效果差异，发现课程学习在稀疏奖励且存在竞争性最优解时几乎是必需的。
该论文提出了一种基于分布差异的马尔可夫决策过程（DMDP）框架，用于学习能稳定状态分布的策略，给出了目标分布渐近稳定的充分条件，并结合PPO和Sac在HalfCheetah任务上验证了分布反馈与显式目标匹配的效果。
AtomSD 提出了一种原子引导的自蒸馏框架，让 LLM 智能体从自身成功轨迹中提取细粒度的"过程原子"（Intent、When、Do、Avoid）并蒸馏进不断演化的策略中，实验显示其平均性能比最先进基线高出最多 9.31 分。
Tsinghua and Didi researchers propose the World Action Critic (WAC) Model, which jointly learns future prediction and consequence evaluation for counterfactual action candidates in world-model-based policy learning, achieving 93.4 PDMS on the NAVSIM-v1 autonomous driving benchmark.
本文提出集体经验学习（CEL）范式，通过失败引导的经验检索微调（FERT）与容错异步训练系统（FATS），让部署中的机器人共享执行经验以改进共同策略，无需额外专家纠正数据。在五项真实操作任务上达到93.3-99.2%的成功率，且更大规模机器人集群可进一步提升性能。
GreenWAM 提出一个机器人世界基础模型，通过共享的解剖学动作空间联合训练 20 个机器人家族，使其既能作为世界模型预测动作后果，又能作为策略控制不同形态的机器人，并在多个基准上取得领先表现。
该论文提出通过学习索引策略（indexical policies）形式的搜索控制，使规划搜索在多项式空间内完成，代价是时间仅随选择深度指数增长；结合语言模型的反例引导学习循环，在IPC 2023学习赛道等基准上解决了1,709/1,890个测试任务，优于LAMA、BFWS和Levitron。
该论文提出CARE，一种针对视觉-语言-动作（VLA）模型推理加速的认证选择框架，通过配对回放和序贯测试，在95%置信度下保证加速-induced失败风险低于用户设定预算，同时在LIBERO基准上实现9.0–10.8倍加速。
Code to Control synthesizes parameterized Python controllers via LLM-generated program structure with derivative-free parameter fitting, enabling real-time control without LLM inference or planning at decision time, and outperforming planning-based program synthesis across Atari, Flappy Bird, and MuJoCo tasks.
The article introduces a browser-based demo for an open-source Clash Royale reinforcement learning environment, where a small REINFORCE policy learns defensive placements with comparisons to brute-force optimal strategies.
This paper introduces Agent Priors-guided Policy Learning (APPL), a framework where construction and runtime agents use structural priors—describing what a behavior depends on—as an interface between skill learning and skill composition, improving out-of-distribution skill generalization on MetaWorld and long-horizon ManiSkill tasks.
The paper introduces CoCA, an on-policy learning framework for dynamic capability allocation in long-horizon multimodal agents, addressing computational overhead and stage-dependent demands through conditional comparisons and reinforcement learning.
MemBodied introduces a fixed-size episodic memory with associative state and episode anchor to enhance vision-language-action models for history-dependent manipulation tasks, achieving significant performance improvements over baselines.
The paper introduces a two-model architecture called Summarize-Judge-Refine (SJR) for multimodal content moderation, which decouples content understanding and policy learning via natural language summaries, enabling significant performance gains and few-shot policy adaptation.
This paper introduces a method to split action generation in vision-language-action models into slow and fast layers, enabling real-time reactive robot control and improving success rates from 42% to 97% with just 10 minutes of online data.
This paper proposes Grounded Action Models (GAMs), a new paradigm for robot foundation models that integrates 3D grounding, achieving state-of-the-art performance on manipulation tasks.
This paper presents a large-scale empirical study of offline reinforcement and imitation learning, analyzing over 160,000 training runs to understand the effects of hyperparameters and dataset properties, and introduces JumpStart, a resource suite for reliable policy-learning research.
Latent Interface Training (LIT) addresses the issue of robot policies learning shortcuts from training images by first teaching action without visual input and then using a pose-supervised latent interface to preserve geometry for action.
PLC-DPO enhances Direct Preference Optimization by routing noisy preference labels into clean, flipped, or tied cases using policy-reference margins, leading to improved performance across various benchmarks.