Thinking Before Looking: Improving Multimodal LLM Reasoning via Mitigating Visual Hallucination

TL;DR

This work addresses hallucinations in multimodal large language models by introducing the Visual Inference Chain (VIC), a thinking-before-looking paradigm that decouples textual reasoning from visual input. VIC uses forward-looking reasoning derived from textual context to generate a VIC, followed by extraction of rationales and a reflective answer inference step, reducing cross-modal biases. Across multiple benchmarks (MMVP, HallusionBench, MME, SEED-Bench, MathVista, POPE) and model families (GPT-4o, Gemini-series), VIC yields robust zero-shot gains and mitigates hallucinations, with ablations showing the importance of VIC generators, single-step extraction, and reflective prompting. The approach demonstrates a practical route to more reliable multimodal reasoning and provides open-source code for replication and further study.

Abstract

Multimodal large language models (MLLMs) have advanced the integration of visual and linguistic modalities, establishing themselves as the dominant paradigm for visual-language tasks. Current approaches like chain of thought (CoT) reasoning have augmented the cognitive capabilities of large language models (LLMs), yet their adaptation to MLLMs is hindered by heightened risks of hallucination in cross-modality comprehension. In this paper, we find that the thinking while looking paradigm in current multimodal CoT approaches--where reasoning chains are generated alongside visual input--fails to mitigate hallucinations caused by misleading images. To address these limitations, we propose the Visual Inference Chain (VIC) framework, a novel approach that constructs reasoning chains using textual context alone before introducing visual input, effectively reducing cross-modal biases and enhancing multimodal reasoning accuracy. Comprehensive evaluations demonstrate that VIC significantly improves zero-shot performance across various vision-related tasks, mitigating hallucinations while refining the reasoning capabilities of MLLMs. Our code repository can be found at https://github.com/Terry-Xu-666/visual_inference_chain.

Figures (13)

• Figure 1: This example from HallusionBench demonstrates the differences between zero-shot, zero-shot CoT, and VIC. The zero-shot CoT represents the thinking while looking approach, which tends to exhibit stereotypical reasoning patterns when processing both visual and textual inputs simultaneously. In contrast, our thinking before looking paradigm, VIC, enhances reasoning quality by decoupling the visual and textual inputs. More examples can be found in Appendix E.
• Figure 2: The overall framework of VIC (Visual Inference Chain). VIC decouples visual and textual inputs to improve reasoning. It first generates intermediate reasoning steps from the question $\mathbf{Q}$ and prompt $\mathbf{P}_{vic}$. The image $\mathbf{I}$ is processed through an MLLM to extract rationales $\mathbf{R}$, which, combined with the visual inference steps $\{s_n\}_{1}^{k}$, lead to the final answer $\mathbf{A}$ with enhanced accuracy.
• Figure 3: Detailed evaluation comparisons with and without VIC for two models on HallusionBench. Features: AA - All Accuracy, HA - Hard Accuracy, EA - Easy Accuracy, FA - Figure Accuracy, QPA - Question Pair Accuracy.
• Figure 4: Performance comparison of Gemini 1.5 Flash and Gemini 1.5 Pro using Zero-shot CoT and VIC methods across various evaluation metrics on HallusionBench.
• Figure 5: Performance of Different VIC Generators. This chart compares the performance of various VIC generators on HallusionBench and SEED-Bench. Grey bars represent the original performance for reference.