Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Understanding the Role of LLMs in Multimodal Evaluation Benchmarks

Botian Jiang, Lei Li, Xiaonan Li, Zhaowei Li, Xiachong Feng, Lingpeng Kong, Qi Liu, Xipeng Qiu

TL;DR

A modified evaluation protocol to disentangle the contributions of the LLM backbone from multimodal integration, and an automatic knowledge identification technique for diagnosing whether LLMs equip the necessary knowledge for corresponding multimodal questions are introduced.

Abstract

The rapid advancement of Multimodal Large Language Models (MLLMs) has been accompanied by the development of various benchmarks to evaluate their capabilities. However, the true nature of these evaluations and the extent to which they assess multimodal reasoning versus merely leveraging the underlying Large Language Model (LLM) backbone remain unclear. This paper presents a comprehensive investigation into the role of LLM backbones in MLLM evaluation, focusing on two critical aspects: the degree to which current benchmarks truly assess multimodal reasoning and the influence of LLM prior knowledge on performance. Specifically, we introduce a modified evaluation protocol to disentangle the contributions of the LLM backbone from multimodal integration, and an automatic knowledge identification technique for diagnosing whether LLMs equip the necessary knowledge for corresponding multimodal questions. Our study encompasses four diverse MLLM benchmarks and eight state-of-the-art MLLMs. Key findings reveal that some benchmarks allow high performance even without visual inputs and up to 50\% of error rates can be attributed to insufficient world knowledge in the LLM backbone, indicating a heavy reliance on language capabilities. To address knowledge deficiencies, we propose a knowledge augmentation pipeline that achieves significant performance gains, with improvements of up to 60\% on certain datasets, resulting in a approximately 4x increase in performance. Our work provides crucial insights into the role of the LLM backbone in MLLMs, and highlights the need for more nuanced benchmarking approaches.

Understanding the Role of LLMs in Multimodal Evaluation Benchmarks

TL;DR

A modified evaluation protocol to disentangle the contributions of the LLM backbone from multimodal integration, and an automatic knowledge identification technique for diagnosing whether LLMs equip the necessary knowledge for corresponding multimodal questions are introduced.

Abstract

The rapid advancement of Multimodal Large Language Models (MLLMs) has been accompanied by the development of various benchmarks to evaluate their capabilities. However, the true nature of these evaluations and the extent to which they assess multimodal reasoning versus merely leveraging the underlying Large Language Model (LLM) backbone remain unclear. This paper presents a comprehensive investigation into the role of LLM backbones in MLLM evaluation, focusing on two critical aspects: the degree to which current benchmarks truly assess multimodal reasoning and the influence of LLM prior knowledge on performance. Specifically, we introduce a modified evaluation protocol to disentangle the contributions of the LLM backbone from multimodal integration, and an automatic knowledge identification technique for diagnosing whether LLMs equip the necessary knowledge for corresponding multimodal questions. Our study encompasses four diverse MLLM benchmarks and eight state-of-the-art MLLMs. Key findings reveal that some benchmarks allow high performance even without visual inputs and up to 50\% of error rates can be attributed to insufficient world knowledge in the LLM backbone, indicating a heavy reliance on language capabilities. To address knowledge deficiencies, we propose a knowledge augmentation pipeline that achieves significant performance gains, with improvements of up to 60\% on certain datasets, resulting in a approximately 4x increase in performance. Our work provides crucial insights into the role of the LLM backbone in MLLMs, and highlights the need for more nuanced benchmarking approaches.

Paper Structure

This paper contains 22 sections, 5 equations, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Problem Notations
  4. Is Visual Capability Necessary?
  5. Do MLLMs Have Sufficient Prior Knowledge?
  6. Can Knowledge Augmentation Improve Multimodal Capabilities?
  7. Experiments
  8. Experimental Setup
  9. LLMs Exploit Shortcuts in Vision Tasks
  10. MLLMs Suffers from LLMs' Knowledge Deficiency
  11. Case Study
  12. Retrieved Knowledge Boosts Multimodal Abilities
  13. Related Work
  14. Multimodal Large Language Models.
  15. Multimodal Understanding Benchmarks.
  16. ...and 7 more sections

Figures (5)

  • Figure 1: Left:We first identify VQA questions answerable without images. Right:We subsequently decompose the process of solving visual questions into two distinct yet interrelated steps, decoupling visual perception capability from knowledge.
  • Figure 2: SR comparison of MLLMs under both image-present and image-absent conditions across four benchmarks. $f(\varnothing, Q)$ is relatively close to $f(I, Q)$, but far from Random Guess, indicating that the model's utilization of visual information is low.
  • Figure 3: SuR and NeR of different models across four datasets. High values indicate that possessing relevant prior knowledge is a prerequisite for solving visual tasks.
  • Figure 4: Top: While successfully answering the visual question, the model fail to perform well on knowledge reasoning tasks. Bottom: The model has relevant knowledge but exhibits hallucinations when addressing visual questions.
  • Figure 5: Differences in SR between scenarios without Retrieval-Augmented Generation (RAG) and those using RAG with 1, 3, 5, and 10 relevant documents. Knowledge enhancement significantly improves model performance.