Reward-Free Evolving Agents Use Pairwise Validators.

Developing self-evolving AI agents often involves a costly process of designing reliable scalar rewards to guide their iterative improvements. These rewards typically require significant domain expertise and expensive labeled examples. A novel approach proposes replacing this scalar reward system with a "pairwise validator." This validator is a frozen Large Language Model (LLM) that, when presented with a parent agent and a child candidate, simply determines which one is superior. This pairwise judgment is generally more stable and easier to obtain than absolute scoring, mitigating the need for strict scale calibration and extensive training. Integrating this pairwise gate into existing self-evolving engines like GEPA, ADRS, and ShinkaEvolve, the method demonstrates performance matching or exceeding full-reward baselines across various agents and artifact types (prompts and code). This offers a competitive task accuracy without the high cost of reward design and labeling.