LLM-as-an-Improver: Using Verification to Generate Better Candidates

Summary

Verifier-based selection usually generates several LLM solutions, ranks them, and discards the verifier’s feedback after choosing a winner. This paper asks whether that feedback can instead improve the candidate pool. Its LLM-as-an-Improver approach, implemented as Verify--Repair--Reselect (VRR), keeps the initial winner and conditionally creates three alternatives: repaired versions of the winner and runner-up, plus a solution using a new approach. Using only inference-time information, VRR removes invalid and duplicate candidates before reselecting an answer with the original evaluation criteria. Across diverse models and code-generation and reasoning benchmarks, VRR improves over fixed-pool verifier-based selection in many settings. It can also recover correct solutions when every candidate in the initial pool is wrong. The results support treating verification as a mechanism for constructing stronger candidates, rather than only ranking existing ones.