Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents
Researchers have developed Co-ReAct, a new framework that uses step-level rubrics to guide ReAct-style AI agents during inference. This approach aims to improve the decision-making process for search-intensive, multi-step reasoning tasks, which often suffer from shallow or redundant trajectories. Co-ReAct injects rubrics into the agent's context at each step to guide evidence seeking, reasoning, and self-evaluation, leading to consistent improvements on benchmarks like DeepResearchBench and SQA-CS-V2. AI
IMPACT Enhances AI agent performance in complex reasoning tasks by providing step-by-step guidance.