Researchers presented the VRR method for improving LLM responses through verification

Researchers have published a paper describing the LLM-as-an-Improver method, which allows models to not just choose the best answer from the proposed ones, but to actively improve their quality. The new approach, called Verify--Repair--Reselect (VRR), uses feedback from a verifier to generate more accurate alternatives during inference.

In contrast to standard methods, where verification is limited to ranking a fixed set of candidates, VRR retains the most promising option and initiates the creation of three additional solutions. The algorithm involves correcting errors in the current leader and the second-best answer, as well as generating a fundamentally new approach. The system automatically filters out invalid and duplicate options, using only data available during inference.

Testing on various code writing and logical thinking benchmarks has shown that the method outperforms traditional strategies for selecting from a fixed pool. VRR is able to recover correct answers even in cases where all initially generated options contained errors. This allows for improving the overall reliability of models without the need for retraining or modifying the architecture of the weights themselves.

The research results emphasize that feedback from a verifier can serve as a tool for constructing stronger candidates, rather than just filtering existing ones. The work is published under the Creative Commons Attribution 4.0 license, which allows the proposed algorithms to be used in third-party developments to improve the accuracy of LLM.