Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents
In brief
- Reward Engine builds success criteria from experience traces to evaluate tool agents.
- Generated rewards guide optimization and improve agent consistency on two benchmarks.
- The method achieves 82.3% and 80.5% consistency with human evaluators on WorkBuddy Bench Office and Claw-Eval.
Open-ended tool agents can achieve goals through multiple paths, but evaluating their quality lacks explicit rewards. Reward Engine automatically generates success criteria from experience traces, then guides agent optimization via these rewards. Results show high consistency with human evaluators on two benchmarks.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error