| # Official Repo of Reagent. |
| Paper: https://arxiv.org/abs/2601.22154 |
|
|
| Code: https://github.com/kxfan2002/Reagent |
|
|
| ## Abstract: |
| Agentic Reinforcement Learning (Agentic RL) has achieved notable success in enabling agents to perform complex reasoning and tool use. |
| However, most methods still relies on sparse outcome-based reward for training. |
| Such feedback fails to differentiate intermediate reasoning quality, leading to suboptimal training results. |
| In this paper, we introduce \textbf{Agent Reasoning Reward Model (Agent-RRM)}, a multi-faceted reward model that produces structured feedback for agentic trajectories, including (1) an explicit reasoning trace , (2) a focused critique that provides refinement guidance by highlighting reasoning flaws, and (3) an overall score that evaluates process performance. |
| Leveraging these signals, we systematically investigate three integration strategies: \textbf{Reagent-C} (text-augmented refinement), \textbf{Reagent-R} (reward-augmented guidance), and \textbf{Reagent-U} (unified feedback integration). |
| Extensive evaluations across 12 diverse benchmarks demonstrate that Reagent-U yields substantial performance leaps, achieving 43.7\% on GAIA and 46.2\% on WebWalkerQA, validating the effectiveness of our reasoning reward model and training schemes. |
|
|
| ## GitHub Repository |
| The official codebase, including training and evaluation scripts for Reagent, can be found on the project's GitHub repository: https://github.com/kxfan2002/Reagent |
|
|
| ## Citation |
|
|
| ```bash |
| @article{fan2026exploring, |
| title={Exploring Reasoning Reward Model for Agents}, |
| author={Fan, Kaixuan and Feng, Kaituo and Zhang, Manyuan and Peng, Tianshuo and Li, Zhixun and Jiang, Yilei and Chen, Shuang and Pei, Peng and Cai, Xunliang and Yue, Xiangyu}, |
| journal={arXiv preprint arXiv:2601.22154}, |
| year={2026} |
| } |
| ``` |
| --- |
| license: apache-2.0 |
| --- |
|
|
