Researchers have a new trick for getting better answers out of large language models: make the quality checker actually fix mistakes instead of just grading them.
The standard approach for boosting LLM accuracy is to generate several candidate answers and use a separate verifier to pick the best one. That verifier's job usually ends at ranking. A new paper introduces Verify-Repair-Reselect (VRR), which instead feeds verification feedback back into the process. VRR keeps the original top candidate but also generates three alternatives: repaired versions of the best and second-best answers, plus one built from a different approach entirely. It then filters out duplicates and invalid options before reselecting a final answer using the same evaluation criteria as before. Tested across multiple models on code-generation and reasoning benchmarks, VRR beat standard fixed-pool selection in many cases.
The notable part is what happens when every initial candidate is wrong. VRR can still recover a correct answer, because the verifier's feedback becomes raw material for a repair rather than a dead end. That reframes verification from a filter into a generator, which matters for anyone building on techniques like best-of-n sampling or self-consistency, where a bad initial batch just means a bad final answer.
It's a small architectural shift, not a new model, but it points at a cheap way to squeeze more reliability out of existing LLMs without retraining them.