Chain-of-Description: What I can understand, I can put into words

TL;DR

This paper introduces Chain-of-Description Prompting, a new method for Multi-Modal Large Language Models that improves performance by encouraging detailed input descriptions before answering, validated on audio and vision benchmarks.

Contribution

The paper proposes Chain-of-Description Prompting, a novel strategy that enhances multi-modal model performance by structured input description, demonstrating significant improvements over standard prompts.

Findings

Nearly 4% improvement on AIR-Bench-Chat audio benchmark

5.3% improvement on MMMU_Pro vision benchmark

Ablation study confirms effectiveness of CoD Prompting

Abstract

In this paper, we propose a novel strategy defined as Chain-of-Description (CoD) Prompting, tailored for Multi-Modal Large Language Models. This approach involves having the model first provide a detailed description of the multi-modal input before generating an answer to the question. When applied to models such as Qwen2-Audio, Qwen2-VL, and Qwen2.5-VL, CoD Prompting significantly enhances performance compared to standard prompting methods. This is demonstrated by nearly a 4\% improvement in the speech category of the audio benchmark AIR-Bench-Chat and a 5.3\% improvement in the hard-level portion of the vision benchmark MMMU\_Pro. Our ablation study further validates the effectiveness of CoD Prompting.