Seeing Through the Chain: Mitigate Hallucination in Multimodal Reasoning Models via CoT Compression and Contrastive Preference Optimization

TL;DR

This work addresses hallucinations in multimodal reasoning models by identifying that reasoning steps can degrade visual grounding and propagate errors into final answers. It introduces C3PO, a two-stage framework combining CoT token compression and contrastive preference optimization, plus a multimodal hallucination-inducing contrast to robustly expose and correct reasoning failures. The authors justify the approach via an Information Bottleneck analysis and show consistent hallucination reduction across diverse MLRMs and benchmarks, while preserving general multimodal capabilities. The results suggest that compact, high-quality reasoning traces, guided by AI feedback and carefully constructed negatives, can markedly enhance reliability in multimodal reasoning systems.

Abstract

While multimodal reasoning models (MLRMs) have exhibited impressive capabilities, they remain prone to hallucinations, and effective solutions are still underexplored. In this paper, we experimentally analyze the hallucination cause and propose C3PO, a training-based mitigation framework comprising \textbf{C}hain-of-Thought \textbf{C}ompression and \textbf{C}ontrastive \textbf{P}reference \textbf{O}ptimization. Firstly, we identify that introducing reasoning mechanisms exacerbates models' reliance on language priors while overlooking visual inputs, which can produce CoTs with reduced visual cues but redundant text tokens. To this end, we propose to selectively filter redundant thinking tokens for a more compact and signal-efficient CoT representation that preserves task-relevant information while suppressing noise. In addition, we observe that the quality of the reasoning trace largely determines whether hallucination emerges in subsequent responses. To leverage this insight, we introduce a reasoning-enhanced preference tuning scheme that constructs training pairs using high-quality AI feedback. We further design a multimodal hallucination-inducing mechanism that elicits models' inherent hallucination patterns via carefully crafted inducers, yielding informative negative signals for contrastive correction. We provide theoretical justification for the effectiveness and demonstrate consistent hallucination reduction across diverse MLRMs and benchmarks.

Figures (11)

• Figure 1: (a) A hallucination case from R1-Onevision yang2025r1. (b) Performance of MLRMs R1-Onevision and MM-Eureka meng2025mm and the base model Qwen2.5-VL-7B bai2025qwen2 on the hallucination benchmark AMBER wang2023llm, where high values indicate fewer hallucinations. The hallucination increases from the base model to reasoning variants.
• Figure 2: Illustration of two important hallucination analyses for MLRMs. (a) Attention distributions of representative MLRMs and the non-reasoning base model Qwen2.5-VL averaged on 300 samples from MSCOCO lin2014microsoft. Compared to Qwen2.5-VL, MLRMs exhibit a substantial reduction in attention to visual tokens and a prominent increase in attention to textual tokens. (b) Proportion of hallucinated answers given hallucinated (-H) and non-hallucinated (-NH) CoTs under CHAIR rohrbach2018object evaluation. EU denotes MM-Eureka and OV denotes R1-Onevision. The hallucination rate of the base model Qwen2.5-VL is reported as a reference.
• Figure 3: Overview of the proposed C3PO framework. We first construct SFT datasets by removing redundant tokens within the generated CoTs based on token importance scores. We then perform hallucination-aware preference optimization, where preferred samples contain enhanced reasoning traces via feedback from an advanced open-source MLLM, while the original unmodified outputs serve as rejected ones. Moreover, we propose a novel multimodal hallucination-inducing mechanism that crafts degraded visual inputs and misleading prompts to elicit MLRM's inherent hallucination patterns as informative negative contrasts.
• Figure 4: GPT-4 assisted benchmark. Sentence-level Hallucination Ratio (SHR) measures the hallucination degree. We also provide 1&2-gram, the number of sentences per image (SPI), and words per image (WPI). A larger radar area indicates better performance.
• Figure 5: Performance on two general-purpose benchmarks.

Theorems & Definitions (6)

• Theorem 1: Compression reduces IB
• Theorem 2: Enhancement reduces IB
• Theorem 1: Compression reduces IB
• proof
• Theorem 2: Enhancement reduces IB
• proof