Verifier Feedback Should Build New Candidates, Not Just Rank the Pool You Already Have

arXiv 2609.19515·medium signalXBlueskyLinkedInCopy link

LLM-as-an-Improver (arXiv 2609.19515, submitted 17 Sep 2026) observes that verifier-based selection discards verification feedback once a fixed candidate pool has been scored. Its Verify-Repair-Reselect loop keeps the initial winner and conditionally generates three complementary alternatives: repaired versions of the winner and the runner-up, plus a solution taking a new approach, then filters invalid and duplicate candidates using inference-time information only and reselects under the original criteria. Across models and code-generation and reasoning benchmarks it beats fixed-pool verifier selection in many settings, and notably can recover a correct solution when every candidate in the initial pool was wrong, which fixed-pool selection cannot do by construction.