Newest papers

One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents

Shaojin Ma, Ruitao Wang, Pengyi Li, Jie Yang, Liangyu Chen, Gangyi Zhao, Hongyao Tang, Yan Zheng, Jianye Hao · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29427 · cited by 0 works (OpenAlex) · open-access PDF

Korean

다중 회전 강화 학습 (RL)의 최근 진보에도 불구하고 복잡하고 장거리 상호 작용 과제에서 LLM 에이전트의 성능을 크게 향상시켰습니다. 그러나, 미묘한 신용 할당의 진전에도 불구하고, 정책 최적화는 응답 특정 확률 비율을 통해 이러한 학습 신호를 적용합니다. 다른 텍스트 응답은 동일한 실행 효과를 유도 할 수 있지만, 후속 결과에서 발생하는 신용 차이점은 그들의 개별 텍스트 실현과 관련이 있습니다. 우리는이 불일치를 신용 업데이트 오류 조정으로 식별합니다. 우리의 분석은 신용 텍스트 쌍이 직접 정책 업데이트 방향을 형성한다는 것을 보여줍니다. 정책 맥락과 실행 효과가 고정되어있을 때조차 EMPO는

Machine translation (Meta M2M100); the English original follows.

Abstract

Recent advances in multi-turn reinforcement learning (RL) have substantially improved the performance of LLM agents on complex, long-horizon interactive tasks. However, despite progress in fine-grained credit assignment, policy optimization still applies these learning signals through response-specific probability ratios. Different textual responses can induce the same executable effect, yet credit differences arising from subsequent outcomes remain tied to their individual textual realizations. We identify this mismatch as credit-update misalignment. Our analysis reveals that credit-text pairing directly shapes policy-update directions, even when the policy context and executable effect are held fixed. Motivated by this finding, we propose Effect-Marginal Policy Optimization (EMPO), which aligns policy optimization with execution-level equivalence through an effect-level surrogate that couples updates across alternative textual realizations. EMPO preserves the original credit signals, integrates with both trajectory-level and step-level credit estimators, and requires no additional critics or rollouts. Experiments on ALFWorld, WebShop, and Sokoban show that EMPO outperforms strong baselines in both task performance and learning efficiency.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Reinforcement learning, Computer science, Executable, Context (archaeology), Task (project management), Artificial intelligence

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange