Debating for Better Reasoning: An Unsupervised Multimodal Approach
Ashutosh Adhikari, Mirella Lapata
TL;DR
This work introduces a multimodal debate framework for visual question answering to enable scalable oversight of vision-language models. Two sighted expert models debate an image-grounded question while a blind, text-only judge decides the winner, focusing on samples where the experts disagree. The authors extend the paradigm to extract reasoning traces from judge verdicts and use them to finetune vision-language models via LoRA, showing that debate generally outperforms individual experts and consultancy. Experiments across MME, MMMU, and MathVista demonstrate improved performance and reasoning capabilities, including in out-of-domain settings, highlighting the practical potential of unsupervised reasoning traces for multimodal AI systems.
Abstract
As Large Language Models (LLMs) gain expertise across diverse domains and modalities, scalable oversight becomes increasingly challenging, particularly when their capabilities may surpass human evaluators. Debate has emerged as a promising mechanism for enabling such oversight. In this work, we extend the debate paradigm to a multimodal setting, exploring its potential for weaker models to supervise and enhance the performance of stronger models. We focus on visual question answering (VQA), where two "sighted" expert vision-language models debate an answer, while a "blind" (text-only) judge adjudicates based solely on the quality of the arguments. In our framework, the experts defend only answers aligned with their beliefs, thereby obviating the need for explicit role-playing and concentrating the debate on instances of expert disagreement. Experiments on several multimodal tasks demonstrate that the debate framework consistently outperforms individual expert models. Moreover, judgments from weaker LLMs can help instill reasoning capabilities in vision-language models through finetuning.