Evolving verifiers improve grading for self-improving AI agents
Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents
Artificial IntelligenceComputation and LanguageMachine Learning
Summary
When AI agents teach themselves new skills, there needs to be a way to check if they are actually improving. The authors created a new kind of grader that evolves alongside the AI agent, learning from its mistakes to become better at judging progress. This evolving grader can spot errors and prevent cheating better than fixed or simple judges. It helps AI systems improve more reliably across tasks like coding, text queries, and report writing.
What this means in practice
- •For software development teams: Use evolving verifiers to reliably evaluate AI-assisted code generation improvements without relying on static benchmarks.
- •For business intelligence teams: Improve text-to-SQL models by co-evolving graders that better detect errors and prevent reward hacking during skill training.
Authors
Xing Zhang, Guanghui Wang, Yanwei Cui, Ziyuan Li, Wei Qiu, Bing Zhu, Peiyang He
Abstract
We changed the agent: did it actually get better? Every self-improving agent loop answers this hundreds of times, and every answer comes from a verifier. On open-ended tasks none exists, so the loop is handed a hand-written rubric or a bare LLM judge grading output from a model like itself, inviting reward hacking and shared blind spots. We make the verifier the evolving object: an inspectable expression over small, mostly deterministic drawback detectors, synthesized from clustered failures, gated at birth, and selected for agreement with a ten-item anchored reference set plus consensus over unlabeled outputs, never for the agent's score. On MBPP+ it gains +0.21 held-out agreement over the hand-authored seed composition, on every seed, and ends ahead of the bare LLM judge it contains. One finding should change how co-evolved verifiers are validated: removing the anchor guards collapses the verifier into a vacuous always-pass grader, yet that collapsed verifier trains skills just as well. Downstream task score cannot certify a self-evolved verifier. Score does answer sufficiency, and there an evolved verifier can substitute: Double Ratchet, pairing the verifier with a lifecycle-managed skill loop, retains 88-110% of the lift that ground truth or a rubric buys the same loop, across code generation, enterprise text-to-SQL, and reference-free report generation. When evolved skills gamed the report rubric, an outer judge caught it and one added detector repaired it; the judge itself was wrong until given the task contract.