Verify by Doing: Co-Evolving Agent Execution and Evaluation without External Feedback
Summary
- Researchers introduce Verify by Doing (VbD), a method that co-evolves Actor and Evaluator harnesses within a single task without external feedback.
- The Evaluator inspects artifacts and environment state using tools to ground internal signals for output selection and procedure updates.
- Proposed Evaluators undergo comparative review, while Actor updates require support from their own fresh outputs.
- VbD outperforms baselines across AutomationBench, tau3 Banking, and GDPval tasks without relying on external feedback.
Why it matters
This research provides a mechanism for AI agents to self-improve and verify success without external feedback, addressing a critical limitation in autonomous deployment scenarios.