EvalResearchBench: Can AI Agents Build Their Own Feedback for Recursive Self-Improvement?

A new benchmark, EvalResearchBench, asks whether AI agents can design their own evaluation feedback to support recursive self-improvement, addressing the cost of repeatedly running complex benchmarks that slows iteration. The work frames the problem around how human experts cut evaluation cost by selecting benchmark subsets or building compact suites, and poses whether AI agents can do the same.

Recursive self-improvement RSI relies on evaluation feedback to assess progress and guide further research, yet repeatedly running complex benchmarks is costly and slows iteration. Human experts reduce this cost by selecting benchmark subsets or designing compact suites. We ask whether AI agents c