Best-of-N methods improve reasoning in text-only large language models by sampling multiple trajectories and selecting an answer. But for vision-language models (VLMs), the paper argues, answer frequency and model confidence alone are not reliable signals.
The arXiv paper introduces VisConf, a quality-aware consensus approach for Best-of-N in VLM reasoning. The abstract does not provide implementation details or experimental results, so the contribution at this stage is the identification of a gap and a proposed direction.
If the method holds up, it could make VLM reasoning more robust in tasks where visual and textual information must be jointly evaluated.