Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Critique Before Thinking: Mitigating Hallucination through Rationale-Augmented Instruction Tuning

Zexian Yang, Dian Li, Dayan Wu, Gang Liu, Weiping Wang

TL;DR

This work tackles visually grounded hallucination in Large Vision-Language Models by introducing Re-Critic, a scalable framework that injects visual rationales into instructions and employs a self-generated preference loop to align outputs. It combines a Visual Rationale Synthesizer (VCIT) with an autoregressive training objective and a Self-Critic Preference Learning pipeline that optimizes via Direct Preference Optimization, avoiding external reward models. Empirically, Re-Critic improves both hallucination-specific benchmarks (POPE, MMHal, HallusionBench, Object HalBench) and general multimodal benchmarks across backbones like LLaVA-v1.5-7B and InternVL2-2B, including notable gains on MathVista and data-efficient scenarios. The results support the claim that methodological learning before reasoning—i.e., embedding rationale context and self-evaluation—enhances context grounding and broad multimodal reasoning capabilities in LVLMs.

Abstract

Despite significant advancements in multimodal reasoning tasks, existing Large Vision-Language Models (LVLMs) are prone to producing visually ungrounded responses when interpreting associated images. In contrast, when humans embark on learning new knowledge, they often rely on a set of fundamental pre-study principles: reviewing outlines to grasp core concepts, summarizing key points to guide their focus and enhance understanding. However, such preparatory actions are notably absent in the current instruction tuning processes. This paper presents Re-Critic, an easily scalable rationale-augmented framework designed to incorporate fundamental rules and chain-of-thought (CoT) as a bridge to enhance reasoning abilities. Specifically, Re-Critic develops a visual rationale synthesizer that scalably augments raw instructions with rationale explanation. To probe more contextually grounded responses, Re-Critic employs an in-context self-critic mechanism to select response pairs for preference tuning. Experiments demonstrate that models fine-tuned with our rationale-augmented dataset yield gains that extend beyond hallucination-specific tasks to broader multimodal reasoning tasks.

Critique Before Thinking: Mitigating Hallucination through Rationale-Augmented Instruction Tuning

TL;DR

This work tackles visually grounded hallucination in Large Vision-Language Models by introducing Re-Critic, a scalable framework that injects visual rationales into instructions and employs a self-generated preference loop to align outputs. It combines a Visual Rationale Synthesizer (VCIT) with an autoregressive training objective and a Self-Critic Preference Learning pipeline that optimizes via Direct Preference Optimization, avoiding external reward models. Empirically, Re-Critic improves both hallucination-specific benchmarks (POPE, MMHal, HallusionBench, Object HalBench) and general multimodal benchmarks across backbones like LLaVA-v1.5-7B and InternVL2-2B, including notable gains on MathVista and data-efficient scenarios. The results support the claim that methodological learning before reasoning—i.e., embedding rationale context and self-evaluation—enhances context grounding and broad multimodal reasoning capabilities in LVLMs.

Abstract

Despite significant advancements in multimodal reasoning tasks, existing Large Vision-Language Models (LVLMs) are prone to producing visually ungrounded responses when interpreting associated images. In contrast, when humans embark on learning new knowledge, they often rely on a set of fundamental pre-study principles: reviewing outlines to grasp core concepts, summarizing key points to guide their focus and enhance understanding. However, such preparatory actions are notably absent in the current instruction tuning processes. This paper presents Re-Critic, an easily scalable rationale-augmented framework designed to incorporate fundamental rules and chain-of-thought (CoT) as a bridge to enhance reasoning abilities. Specifically, Re-Critic develops a visual rationale synthesizer that scalably augments raw instructions with rationale explanation. To probe more contextually grounded responses, Re-Critic employs an in-context self-critic mechanism to select response pairs for preference tuning. Experiments demonstrate that models fine-tuned with our rationale-augmented dataset yield gains that extend beyond hallucination-specific tasks to broader multimodal reasoning tasks.
Paper Structure (11 sections, 3 equations, 5 figures, 5 tables)

This paper contains 11 sections, 3 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Visual Rationale Synthesizer
  5. Self-Critic Prefernece Learning
  6. Experiments
  7. Experimental Setup
  8. Main Results
  9. Ablation Study
  10. Further Analysis
  11. Conclusion

Figures (5)

  • Figure 1: Comparison of hallucination mitigation methods. In this example, our method provides the rationale explanation prior to reasoning, enabling more accurate interpretation of an object's color by considering its reflective properties.
  • Figure 2: Overview of Re-Critic. The overall process is divided into three parts: 1) augmented standard QA with synthesized visual rationales, then 2) fine-tuning LVLM with rewritten QA, and finally 3) perform DPO through the self-critic preference learning.
  • Figure 3: Ablation study on the effectiveness of using the alternative base model MiniGPT4 (left) and smaller scale of training data (right).
  • Figure 4: Performance comparison of different models on six general tasks and one hallucination task.
  • Figure 5: A case study of different predictions with LLaVA 1.5 as the backbone model.