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

Indonesian

Kemajuan baru-baru ini dalam pembelajaran multi-turn reinforcement (RL) telah secara substansial meningkatkan kinerja agen LLM pada tugas-tugas interaktif yang kompleks, jangka panjang. Namun, terlepas dari kemajuan dalam penyerahan kredit yang efisien, optimasi kebijakan masih menerapkan sinyal pembelajaran ini melalui rasio probabilitas respon spesifik. Tanggapan teks yang berbeda dapat menginduksi efek eksekusi yang sama, namun perbedaan kredit yang timbul dari hasil-hasil berikutnya tetap terkait dengan realisasi teks individu mereka. Kami mengidentifikasi ketidaksamaan ini sebagai misalignment kredit-update. Analisis kami mengungkapkan bahwa pasang kredit-tekst secara langsung membentuk arah pembaruan kebijakan, bahkan ketika konteks kebijakan dan efek eksekusi tetap tetap tetap tetap. Bermotivasi oleh temuan ini, kami mengusulkan Efek-Margin Policy Optimization

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