One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents
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.
Related
- Closing the Horizon Gap in Policy Optimization for Adversarial MDPs
- Dual-Axis Policy Optimization for LLM Agents: Bayesian Feedback Attribution and Trajectory Mass Normalization
- Benign Nonconvex Landscape for Policy Optimization: Infinite-Horizon Discounted MDPs with General State and Action Spaces
- Bellman Policy Optimization
- Proximal Policy Optimization for Amortized Discrete Sampling
BibTeX
@misc{ma2026many,
title = {{One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents}},
author = {Shaojin Ma and Ruitao Wang and Pengyi Li and Jie Yang and Liangyu Chen and Gangyi Zhao and Hongyao Tang and Yan Zheng and Jianye Hao},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.29427},
eprint = {2610.29427},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.29427},
}
RIS
TY - UNPB
TI - One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents
AU - Shaojin Ma
AU - Ruitao Wang
AU - Pengyi Li
AU - Jie Yang
AU - Liangyu Chen
AU - Gangyi Zhao
AU - Hongyao Tang
AU - Yan Zheng
AU - Jianye Hao
PY - 2026
DA - 2026/10/21
PB - arXiv
DO - 10.48550/arxiv.2610.29427
UR - https://arxiv.org/abs/2610.29427
AB - 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.
ER -