Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data

Yucheng Shi, Quanzheng Li, Jin Sun, Xiang Li, Ninghao Liu

TL;DR

This work tackles the gap in domain-specific visual cognition and explainability in large multimodal models by introducing a self-synthesized data framework. It combines an Information Bottleneck-based image-level concept selection with a reward-model-free rejection sampling mechanism to iteratively fine-tune LMMs on interpretable, image-grounded explanations. Empirical results across diverse fine-grained and medical datasets show improved classification accuracy and high-quality explanations, with explanations consistently produced (EE = 1.00) and cognitive coherence enhanced (CS) while maintaining fluency. The approach achieves tangible improvements without heavy manual annotations and offers robust visualization and usability benefits for specialized applications.

Abstract

Large Multimodal Models (LMMs), or Vision-Language Models (VLMs), have shown impressive capabilities in a wide range of visual tasks. However, they often struggle with fine-grained visual reasoning, failing to identify domain-specific objectives and provide justifiable explanations for their predictions. To address the above challenge, we propose a novel visual rejection sampling framework to improve the cognition and explainability of LMMs using self-synthesized data. Specifically, visual fine-tuning requires images, queries, and target answers. Our approach begins by synthesizing interpretable answers that include human-verifiable visual features. These features are based on expert-defined concepts, and carefully selected based on their alignment with the image content. After each round of fine-tuning, we apply a reward model-free filtering mechanism to select the highest-quality interpretable answers for the next round of tuning. This iterative process of synthetic data generation and fine-tuning progressively improves the model's ability to generate accurate and reasonable explanations. Experimental results demonstrate the effectiveness of our method in improving both the accuracy and explainability of specialized visual classification tasks.

Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data

TL;DR

This work tackles the gap in domain-specific visual cognition and explainability in large multimodal models by introducing a self-synthesized data framework. It combines an Information Bottleneck-based image-level concept selection with a reward-model-free rejection sampling mechanism to iteratively fine-tune LMMs on interpretable, image-grounded explanations. Empirical results across diverse fine-grained and medical datasets show improved classification accuracy and high-quality explanations, with explanations consistently produced (EE = 1.00) and cognitive coherence enhanced (CS) while maintaining fluency. The approach achieves tangible improvements without heavy manual annotations and offers robust visualization and usability benefits for specialized applications.

Abstract

Large Multimodal Models (LMMs), or Vision-Language Models (VLMs), have shown impressive capabilities in a wide range of visual tasks. However, they often struggle with fine-grained visual reasoning, failing to identify domain-specific objectives and provide justifiable explanations for their predictions. To address the above challenge, we propose a novel visual rejection sampling framework to improve the cognition and explainability of LMMs using self-synthesized data. Specifically, visual fine-tuning requires images, queries, and target answers. Our approach begins by synthesizing interpretable answers that include human-verifiable visual features. These features are based on expert-defined concepts, and carefully selected based on their alignment with the image content. After each round of fine-tuning, we apply a reward model-free filtering mechanism to select the highest-quality interpretable answers for the next round of tuning. This iterative process of synthetic data generation and fine-tuning progressively improves the model's ability to generate accurate and reasonable explanations. Experimental results demonstrate the effectiveness of our method in improving both the accuracy and explainability of specialized visual classification tasks.

Paper Structure

This paper contains 34 sections, 4 theorems, 16 equations, 8 figures, 15 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Methodology
  4. Overview
  5. Step 1: Image-level Visual Concept Selection
  6. Step 2: Reward Model-Free Rejection Sampling
  7. Theoretical Justification
  8. Experiments
  9. Experiments Settings
  10. RQ1: Training on our synthetic data improves classification
  11. RQ2: Our method provides high-quality explanation
  12. RQ3: Evaluation on Interpretable Concepts Selection
  13. RQ4: Visualization of Visual Focus of LMMs
  14. RQ5: Evaluation on Usability
  15. Related Work: Self-improvement and Data Synthesis
  16. ...and 19 more sections

Key Result

Theorem 1

Let $X$ be the true image content with label $c$ and $D = \{d_1, d_2, \dots, d_n\}$ be independent and identically distributed (i.i.d.) samples from $P(D|X)$. Let $Z$ be an expert-defined concept list about label $c$. Under the assumptions of conditional independence and convergence (Assumptions ass

Figures (8)

  • Figure 1: LLaVA-1.5 struggles to utilize key visual features in images for reasoning and explaining predictions in classification tasks.
  • Figure 2: Examples of synthetic answers for query Q. Training with the first two types leads to shortcut learning or overgeneralization.
  • Figure 3: Our framework: An iterative approach of data synthesis and model fine-tuning.
  • Figure 4: Our generated answers contain detailed visual explanations.
  • Figure 5: Our method demonstrates superior precision in concept selection compared to applying GPT-4o.
  • ...and 3 more figures

Theorems & Definitions (6)

  • Theorem 1
  • Theorem 2
  • Theorem 3
  • proof
  • Theorem 4
  • proof