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

Spanish

Los avances recientes en el aprendizaje de reforzamiento multi-turn (RL) han mejorado sustancialmente el rendimiento de los agentes de LLM en tareas interactivas complejas y de largo horizonte. Sin embargo, a pesar del progreso en la asignación de crédito de grano fino, la optimización de políticas sigue aplicando estas señales de aprendizaje a través de relaciones de probabilidad específicas de respuesta. Las diferentes respuestas textuales pueden inducir el mismo efecto ejecutable, pero las diferencias de crédito derivadas de los resultados posteriores siguen estando vinculadas a sus realizaciones textuales individuales. Identificamos esta desajustación como equivalencia de actualización de crédito. Nuestro análisis revela que el apareamiento de texto a crédito forma directamente direcciones de actualización de políticas, incluso cuando el contexto de políticas y el efecto ejec

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