Method repairs wrong AI answers before picking a winner

Verifier-based selection wastes its own feedback the moment it finishes ranking candidates, according to a paper on Verify-Repair-Reselect posted to arXiv.

That matters for anyone building agent workflows around sample-then-verify: coding agents, math solvers, tool-calling loops that generate a few attempts and pick the best one. The standard approach scores a fixed batch and keeps the winner. Once the pool is scored, nothing that happens next changes it, and if every candidate in the batch is wrong, the system just picks the least-wrong loser.

Verify-Repair-Reselect changes what happens after scoring. It keeps the current winner in play. Then it conditionally generates up to three new candidates: a repaired version of the winner, a repaired version of the runner-up, and a solution built from a different approach. Invalid or duplicate candidates get filtered using only information available at inference time.

The pool grows and improves instead of getting scored once and discarded. The paper reports Verify-Repair-Reselect beats fixed-pool selection in many of the settings tested. The sharper claim is what happens when every original candidate is wrong: it can still generate and select a correct answer, something ranking alone can't do by construction.

The paper doesn't say how the repair step's cost scales as candidate pools grow, or how much extra inference it adds on top of the original generation calls. That's the number that decides whether this is a drop-in swap for existing verifier setups or a meaningfully pricier one.