Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents
The paper introduces Co-ReAct, a framework that enhances ReAct agents by using rubrics as step-level guidance during multi-step reasoning tasks. This approach aims to improve the decision-making process of agents by providing targeted evidence-seeking and reasoning strategies. The results show that Co-ReAct consistently outperforms existing methods across various benchmarks.
- ▪Co-ReAct uses rubrics to guide agents in their decision-making process during inference.
- ▪The framework was tested on DeepResearchBench and SQA-CS-V2, showing significant improvements over traditional ReAct methods.
- ▪A dedicated rubric generator was trained to optimize the guidance provided to agents.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.23590 (cs) [Submitted on 22 May 2026] Title:Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents Authors:Jiazheng Kang, Bowen Zhang, Zixin Song, Jiangwang Chen, Xiao Yang, Da Zhu, Guanjun Jiang View a PDF of the paper titled Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents, by Jiazheng Kang and 6 other authors View PDF HTML (experimental) Abstract:ReAct-style agents for search-intensive, multi-step reasoning tasks rely largely on their own internal judgment to decide what evidence to seek, which reasoning or action step to take next, and when to stop, often producing shallow, redundant, or poorly targeted trajectories.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
