The paper introduces the Implicit Prefix-Value Reward Model (IPVRM), which learns the probability of eventual correctness for each prefix directly from outcome labels, thereby aligning training targets with inference-time step signals via temporal-difference differences. IPVRM improves step-verification F1 on ProcessBench. Additionally, the authors propose Distribution-Level RL (DistRL), a policy optimization method that applies TD advantages to both sampled and high-probability tokens, offering dense counterfactual updates without extra rollouts, and show that DistRL consistently enhances downstream reasoning when combined with IPVRM.
By Shiping Gao, Hongzhan Chen, Xiaojun Quan, Qifan Wang, Lifu Huang
arXiv:2607. 03702v1 Announce Type: new Abstract: Large language model (LLM) agents have shown strong decision-making capabilities in long-horizon interactive tasks, yet they still struggle to effectively leverage failed trajectories: full retries incur high interaction costs, while experience retrieval tends to dilute critical experience signals.
By Weiyang Guo, Zesheng Shi, Longhui Zhang, Zeen Zhu, Min Zhang, Jing Li
The paper introduces Potential-Guided Policy Optimization (PGPO), a method for multi-turn agentic tasks that improves credit assignment by estimating empirical state potentials from anchor-state-group return statistics. PGPO derives action advantages from potential differences between adjacent states, enabling cross-trajectory credit propagation and finer-grained step-level credit assignment, especially within failed trajectories. Experiments on ALFWorld and WebShop demonstrate strong performance compared to recent group-based reinforcement learning methods, with negligible training overhead.
By Yuyao Zheng, Haipeng Sun, Junwei Bao, Lemao Liu, Hongfei Jiang, Yang Song, Dejing Dou
arXiv:2607. 22724v1 Announce Type: cross Abstract: Group-based policy optimization has been increasingly used to train large language model (LLM) agents from sparse outcome rewards by comparing trajectories or steps within a group.
By Kaibing Yang, Guangfeng Cai, Shengtian Yang, Shuo He, Yu Li, Mengyi Liu, Pengwei Chen, Jun Xu, Lei Feng
arXiv:2608. 05102v1 Announce Type: new Abstract: Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer.
By Yijun Lu, Rui Ye, Jiajun Wang, Yuwen Du, Tian Jin, Songhua Liu, Siheng Chen
arXiv:2608. 12764v1 Announce Type: cross Abstract: Deep search agents operate over trajectories spanning dozens of steps, yet standard reinforcement learning provides only a single outcome reward per trajectory, which is far too sparse for effective credit assignment.
By Haoze Wu, Chuqiao Kuang, Tianyi Zhuang, Xiaoguang Li
arXiv:2608. 07959v1 Announce Type: new Abstract: Ultra-long egocentric video understanding requires reasoning over temporally sparse evidence distributed across hours or days, challenging current multimodal models with limited context and the grounding of key video segments.
By Keyang Zhong, Kuo Wang, Peng Liu, Quanlong Zheng, Junlin Xie, Zhijia Liang, Yanhao Zhang, Guanbin Li
The paper explores how dense, turn-level reward structures can improve reinforcement learning for large language model agents in multi-turn tasks. It introduces three reward granularity types—terminal, delayed, and per-turn—and adapts Group Relative Policy Optimization and Proximal Policy Optimization to each. Experiments on search and game agents show that per-turn rewards consistently yield better training dynamics, faster convergence, and higher answer correctness compared to sparse terminal or delayed rewards.
By Quan Wei, Siliang Zeng, Chenliang Li, Zhongruo Wang, William Brown, Oana Frunza, Wei Deng, Anderson Schneider, Yuriy Nevmyvaka, Yang Katie Zhao, Alfredo Garcia, Mingyi Hong
arXiv:2602. 05459v2 Announce Type: replace Abstract: Offline goal-conditioned reinforcement learning (GCRL) is typically benchmarked by the best tuned success rate of each method.
By Jan Malte T\"opperwien, Aditya Mohan, Marius Lindauer
The paper introduces GRAFT, a Graph-based Faithful sTep-level credit-assignment framework that constructs a trajectory graph from rollout trajectories, recovers node state-values via Bellman iteration, and assigns step-level advantages based on node value differences. It also proposes Graph GAE to further reduce state-value estimation bias. Experiments on multi-turn agentic benchmarks demonstrate consistent improvements over GRPO and other recent agentic RL algorithms.
By Xincheng Yao, Haobo Fu, Weiming Liu, Chongyang Zhang
arXiv:2608. 03223v1 Announce Type: cross Abstract: Agentic reinforcement learning enables LLM agents to learn through interaction, but sparse trajectory-level rewards reveal success without identifying which intermediate decisions deserve credit.
By Ranxu Zhang, Guinan Chen, Chenshaodong, Jinghao Lin, Xiaozhou Xu, Sunzhe, Yanyong Zhang, Chao Wang
Large language models (LLMs) are increasingly extended into deep search agents that solve complex questions through multi-step interaction with external search and browsing tools. However, existing agents often incur substantial computational and interaction costs, generating lengthy trajectories that contain redundant queries, inefficient exploration, and irrelevant observations.