Debating for Better Reasoning: An Unsupervised Multimodal Approach

TL;DR

This paper introduces a multimodal debate framework where vision-language models debate answers, and a text-only judge evaluates them, leading to improved performance and reasoning in models, especially for visual question answering tasks.

Contribution

It extends the debate paradigm to multimodal settings, enabling weaker models to supervise and enhance stronger models' reasoning capabilities.

Findings

Debate framework outperforms individual models on multimodal tasks.

Weaker LLM judgments can improve vision-language model reasoning.

The approach reduces reliance on explicit role-playing in debates.

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.