V-STaR: Training Verifiers for Self-Taught Reasoners
This addresses a bottleneck in self-improvement for LLMs, offering a more efficient method for enhancing reasoning and code generation tasks, though it is incremental as it builds on existing self-improvement frameworks.
The paper tackles the problem of discarding incorrect solutions in self-improvement methods for large language models by proposing V-STaR, which trains a verifier using both correct and incorrect solutions to select the best candidate at inference, resulting in a 4% to 17% test accuracy improvement over existing approaches on code generation and math reasoning benchmarks.
Common self-improvement approaches for large language models (LLMs), such as STaR, iteratively fine-tune LLMs on self-generated solutions to improve their problem-solving ability. However, these approaches discard the large amounts of incorrect solutions generated during this process, potentially neglecting valuable information in such solutions. To address this shortcoming, we propose V-STaR that utilizes both the correct and incorrect solutions generated during the self-improvement process to train a verifier using DPO that judges correctness of model-generated solutions. This verifier is used at inference time to select one solution among many candidate solutions. Running V-STaR for multiple iterations results in progressively better reasoners and verifiers, delivering a 4% to 17% test accuracy improvement over existing self-improvement and verification approaches on common code generation and math reasoning benchmarks with LLaMA2 models.