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 Chain-of-Thought Compression and Contrastive Preference Optimization. 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 & tables
Figure 1: (a) A hallucination case from R1-Onevision Yang et al. (2025a) . (b) Performance of MLRMs R1-Onevision and MM-Eureka Meng et al. (2025) , and the base model Qwen2.5-VL-7B Bai et al. (2025b) on AMBER Wang et al. (2023) , where high values indicate fewer hallucinations . The hallucination increases from the base model to reasoning variants.
Figure 2: Two important hallucination analyses for MLRMs. (a) Attention distributions of MLRMs and the non-reasoning Qwen2.5-VL on 300 samples from MSCOCO Lin et al. (2014) . As observed, MLRMs exhibit a substantial reduction in attention to visual tokens. (b) Proportion of hallucinated answers given hallucinated (-H) and non-hallucinated (-NH) CoTs under CHAIR Rohrbach et al. (2018) evaluation. EU denotes MM-Eureka, and OV denotes R1-Onevision. The result of Qwen2.5-VL is reported as a reference.
Figure 3: Overview of the proposed C3PO framework. We first construct SFT datasets by removing redundant low-importance tokens in generated CoTs. Next, we perform hallucination-aware preference optimization, where preferred samples contain CoTs enhanced by an advanced open-source MLLM, while the original unmodified outputs serve as rejected ones. To further enhance the optimization, we propose a multimodal hallucination-inducing technique that crafts degraded visual inputs and misleading prompts to elicit MLRM’s inherent hallucination patterns as negative contrasts.
Method
R1-Onevision
Orsta-R1
MM-Eureka
MM-R1
ThinkLite
CS↓
CI↓
CS↓
CI↓
CS↓
CI↓
CS↓
CI↓
CS↓
CI↓
Vanilla
36.7
8.2
40.7
8.5
44.0
8.6
37.0
7.4
41.7
7.8
RLAIF-V
39.0
8.3
39.7
7.6
41.3
8.5
38.7
7.5
40.3
7.5
OPA-DPO
36.7
7.7
41.7
8.2
41.3
7.7
36.3
7.2
40.0
7.9
C3PO
35.0
7.4
35.3
6.9
40.0
7.7
35.0
7.2
38.0
7.3
Table 1: Comparison of the proposed C3PO with competitive baselines on the CHAIR benchmark. We evaluate on images from MSCOCO. CS and CI denotes CHAIR S and CHAIR I respectively.
Model
Method
Random
Popular
Adversarial
Acc.
F1 score
Acc.
F1 score
Acc.
F1 score
MM-Eureka
Vanilla
84.12
81.90
83.90
81.23
83.47
80.85
RLAIF-V
83.81
81.46
83.70
80.89
83.27
80.47
OPA-DPO
85.29
83.46
84.93
82.70
84.37
82.16
C3PO
86.22
84.68
85.87
83.94
85.23
83.34
MM-R1
Vanilla
84.85
82.83
84.47
82.04
83.97
81.56
Table 2: Comparison of the proposed C3PO with competitive baselines on the POPE benchmark (%). We report the accuracy and F1 score for evaluation.
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.
Method
R1-Onevision
Orsta-R1
MM-Eureka
MM-R1
ThinkLite
Acc.
F1 score
Acc.
F1 score
Acc.
F1 score
Acc.
F1 score
Acc.
F1 score
Vanilla
81.52
86.54
82.94
87.31
83.27
87.15
84.38
88.20
82.25
86.96
RLAIF-V
80.99
84.97
83.14
87.50
83.46
87.49
84.17
88.11
82.31
87.04
OPA-DPO
80.71
85.30
83.58
87.67
83.72
87.59
84.76
88.51
83.02
87.36
C3PO
81.90
86.73
84.57
88.50
85.01
88.80
85.61
89.25
83.96
88.26
Table 3: Comparison of the proposed C3PO with competitive baselines on the AMBER benchmark (%). We report the accuracy and F1 score for evaluation.
Figure 8
Method
CHAIR Eval
GPT-4 Eval
CS↓
CI↓
SHR ↓
C3PO
35.3
6.9
26.3
w/o loss LRE
37.3
7.2
28.8
w/o loss LDPO
39.7
8.1
29.4
w/o loss LAnc
37.7
7.5
27.2
w/o MHI strategy
37.0
7.9
28.4
Table 4: Ablation study of the proposed techniques in C3PO under CHAIR and GPT-4 assisted evaluation benchmarks.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Method
Random
Popular
Adversarial
Accuracy
F1 score
Accuracy
F1 score
Accuracy
F1 score
R1-Onevision
Vanilla
83.64
81.82
82.80
80.59
81.93
79.81
RLAIF-V
82.85
81.03
81.60
79.45
81.57
79.42
OPA-DPO
83.85
82.28
82.57
80.67
82.03
80.19
C3PO
83.88
82.66
82.53
81.01
82.37
80.87
Orsta-R1
Vanilla
85.15
83.29
84.70
82.43
84.13
81.90
Appendix
Table 5: Additional performance comparison of the proposed C3PO and strong baselines on the POPE benchmark. We supplement results on MLRMs Orsta-R1 and R1-Onevision.
Figure 7: GPT-4 assisted benchmark results for R1-Onevision. A larger radar area indicates better general performance.
Metric
R1-Onevision
Orsta-R1
MM-Eureka
MM-R1
ThinkLite
C3PO
w/o DPO
Vanilla
C3PO
w/o DPO
Vanilla
C3PO
w/o DPO
Vanilla
C3PO
w/o DPO
Vanilla
C3PO
w/o DPO
Vanilla
CS↓
35.0
40.0
36.7
35.3
39.7
40.7
40.0
40.7
44.0
35.0
35.3
37.0
38.0
41.3
41.7
CI↓
7.4
7.7
8.2
6.9
7.7
8.5
7.7
8.3
8.6
7.2
6.9
7.4
7.3
7.2
7.8
SHR ↓
37.9
36.7
40.7
26.3
29.2
30.6
28.1
29.3
30.6
27.9
31.0
29.8
28.5
29.9
32.1
Appendix
Table 6: Ablation study of C3PO on the preference optimization and the CoT compression.
Figure 8: Ablation study on the weight factor λDPO of the LDPO terms.
Figure 9: Visualization of the pruned CoTs.
Figure 10: Visualization results comparing our C3PO and other methods. Hallucinations are marked in red .
Figure 11: Visualization results comparing our C3PO and other methods. Hallucinations are marked in red .
Multimodal Large Reasoning Models introduce the reasoning paradigm, demonstrating strong capabilities on complex vision-language tasks. However, they still suffer from severe hallucinations. Existing training-based methods typically mitigate hallucinations through response-level direct preference optimization (DPO), where the Chain-of-Thought (CoT) and the final answer are treated as a monolithic output and optimized jointly. We reveal that this formulation performs similarly to answer-only optimization, suggesting that it primarily learns answer-level preference, while leaving CoT-level supervision insufficiently exploited. To address this issue, we explicitly formulate a CoT-oriented preference term and derive Reasoning-Conditioned Direct Preference Optimization (RC-DPO), which models the CoT as a condition for answer generation and contrasts the preference for the same preferred answer under different CoT conditions, promoting answer-supportive reasoning chain alignment. To further improve optimization, we introduce a reasoning-enhanced preference data generation strategy that employs Monte Carlo Tree Search to discover visually grounded and logically consistent CoTs as positive samples, and attention-guided CoT token pruning to construct negative ones. Extensive experiments across various models and benchmarks show that RC-DPO effectively mitigates hallucinations and improves the reliability of the multimodal reasoning process.
Jiawei Kong, Hao Fang, Shunxiang Liao +5
Tsinghua Shenzhen International Graduate School, Tsinghua University · Harbin Institute of Technology, Shenzhen
Multimodal Chain-of-Thought (MCoT) models have demonstrated impressive capability in complex visual reasoning tasks. Unfortunately, recent studies reveal that they suffer from severe hallucination problems due to diminished visual attention during the generation process. However, visual attention decay is a well-studied problem in Large Vision-Language Models (LVLMs). Considering the fundamental differences in reasoning processes between MCoT models and traditional LVLMs, we raise a basic question: Whether MCoT models have unique causes of hallucinations? To answer this question, we systematically investigate the hallucination patterns of MCoT models and find that fabricated texts are primarily generated in associative reasoning steps, which we term divergent thinking. Leveraging these insights, we introduce a simple yet effective strategy that can effectively localize divergent thinking steps and intervene in the decoding process to mitigate hallucinations. Extensive experiments show that our method outperforms existing methods by a large margin. More importantly, our proposed method can be conveniently integrated with other hallucination mitigation methods and further boost their performance. The code is publicly available at https://github.com/ASGO-MM/MCoT-hallucination.
Ji Ma, Wei Suo, Peng Wang +1
School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, China. · National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, China.
Multimodal Large Reasoning Models (MLRMs) have achieved remarkable strides in visual reasoning through test time compute scaling, yet long chain reasoning remains prone to hallucinations. We identify a concerning phenomenon termed the Reasoning Vision Truth Disconnect (RVTD): hallucinations are strongly correlated with cognitive bifurcation points that often exhibit high entropy states. We attribute this vulnerability to a breakdown in visual semantic anchoring, localized within the network's intermediate layers; specifically, during these high uncertainty transitions, the model fails to query visual evidence, reverting instead to language priors. Consequently, we advocate a shift from solely outcome level supervision to augmenting it with fine grained internal attention guidance. To this end, we propose V-STAR (Visual Structural Training with Attention Reinforcement), a lightweight, holistic training paradigm designed to internalize visually aware reasoning capabilities. Central to our approach is the Hierarchical Visual Attention Reward (HVAR), integrated within the GRPO framework. Upon detecting high entropy states, this mechanism dynamically incentivizes visual attention across critical intermediate layers, thereby anchoring the reasoning process back to the visual input. Furthermore, we introduce the Forced Reflection Mechanism (FRM), a trajectory editing strategy that disrupts cognitive inertia by triggering reflection around high entropy cognitive bifurcation points and encouraging verification of subsequent steps against the visual input, thereby translating external debiasing interventions into an intrinsic capability for hallucination mitigation.
Zhe Qian, Yanbiao Ma, Zhuohan Ouyang +7
South China Agricultural University · Gaoling School of Artificial Intelligence, Renmin University of China · South China Normal University +4