AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Seeing Through the Chain: Understanding and Mitigating Hallucinations in Multimodal Large Reasoning Models

Hao Fang, Jinyu Li, Jiawei Kong, Tianqu Zhuang, Kuofeng Gao, Bin Chen, Shu-Tao Xia

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Investigates why multimodal large reasoning models hallucinate more and proposes a reasoning-centric training framework to mitigate hallucinations.

Problem

Multimodal large reasoning models (MLRMs) remain prone to hallucinations, and effective mitigation methods are underexplored.

Approach

Attributes MLRM hallucinations to reasoning mechanisms that increase reliance on language priors, and proposes a reasoning-centric training framework to counter this.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add C3PO (CoT compression + contrastive preference optimization)Orsta-R1-7B · CHAIRCHAIR_S: 40.7 -> 35.3Improved
apply RLAIF-V baseline to MLRMR1-Onevision-7B · CHAIRCHAIR_S: 36.7 -> 39.0Hurt
apply OPA-DPO baseline to MLRMOrsta-R1-7B · CHAIRCHAIR_S: 40.7 -> 41.7Hurt
add C3POMM-Eureka-7B · POPE Randomaccuracy (%): 84.12 -> 86.22Improved
add C3POMM-Eureka-7B · AMBER discriminativeaccuracy (%): 83.27 -> 85.01Improved
add reasoning-enhanced preference loss L_RE (ablation)C3PO · CHAIRCHAIR_S: 37.3 -> 35.3Improved
add standard DPO loss L_DPO over joint output (ablation)C3PO · CHAIRCHAIR_S: 39.7 -> 35.3Improved
add anchor regularization L_Anc (ablation)C3PO · CHAIRCHAIR_S: 37.7 -> 35.3Improved
add multimodal hallucination-inducing negatives (ablation)C3PO · GPT-4 assisted evalSHR: 28.4 -> 26.3Improved
add preference optimization after SFT on compressed CoT (ablation)Orsta-R1-7B · CHAIR / GPT-4 evalCHAIR_S and SHR: reductions of 4.4 and 2.9 in C_S and SHRImproved
SFT on compressed CoT alone (no DPO) (ablation)R1-Onevision-7B · GPT-4 assisted evalSHR: w/o DPO 36.7 vs C3PO 37.9Mixed
prune CoT tokens (gamma < 1.0) instead of no pruning (ablation)Orsta-R1-7B · CHAIR and GPT-4 evalCHAIR / SHRImproved

Surprising findings

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy