Current multimodal large language models (MLLMs) suffer from unstable reasoning due to the lack of mechanisms for controlling reasoning trajectories, often leading to hallucinated conclusions. Our analysis reveals that early-stage uncertainty induces trajectory switching, significantly increasing hallucination risk. We propose MedMMV, an evidence-grounded trajectory control framework that treats reasoning as a search-and-selection process over candidate trajectories. By constructing, refining, and selecting among multiple reasoning paths, MedMMV improves reliability and produces structured supervision signals over trajectory quality. We further convert these signals into learnable objectives for post-training, including reject sampling and pairwise preference learning, enabling smaller models to learn trajectory selection as an efficient single-pass capability. Instead of merely imitating reasoning traces, the model is trained to assign higher likelihood to preferred trajectories. Experiments show that richer supervision leads to stronger transfer, with accuracy improving from 10.5% to 29.0%. Our results demonstrate that trajectory selection is the key transferable capability for reliable reasoning.
Figures & tables
Figure 1: Overview of MedMMV for trajectory-level control and signal-driven learning in multimodal clinical reasoning. Top-left : A multimodal patient case. Top-right : The student model trained with trajectory-level supervision significantly outperforms trace imitation and approaches the performance of full trajectory control. Bottom-left : Standard chain-of-thought reasoning suffers from early uncertainty, leading to trajectory divergence. Bottom-middle : MedMMV performs trajectory-level control through three stages. Bottom-right : The MedMMV process produces structured trajectory-level signals, which we transfer to a smaller model through post-training.
System
Data Modality
Core Reasoning Engine
Structured Reasoning & Validation
Text
Image
Self- Revision
Evidence Grounding
Halluc. Control
Parallel Exploration
Active Supervision
Consensus & Aggregation
Traceable Reasoning Path
Reasoning Methods
CoT ( Wei et al., 2023 )
✓
✗
✗
✗
✗
✗
✗
✗
✗
MedPaLM ( Singhal et al., 2023 )
✓
✓ ∗
✓
✓
✓
✗
✗
✓ ∗∗
✗
Multi-Agent Systems
ClinicalAgent ( Yue et al., 2024 )
✓
✗
✓
✗
✗
✗
✗
✗
✗
Table 1: Comparison of MedMMV with representative prior methods.
Figure 2: Overview of MedMMV as a trajectory-level control and amortized selection framework. Left : MedMMV performs inference-time trajectory control via proposal, evidence-grounded repair, and CU-based selection. Right : The same pipeline produces trajectory-level signals that are converted into supervision through reject sampling, preference learning, and on-policy reranking. By optimizing likelihood over preferred trajectories, the student model amortizes trajectory selection into single-pass generation.
Table 2: Overall accuracy results on six medical benchmarks. T : text modality; I : image modality. Red shading highlights the top three methods for each dataset.
Table 3: Truthfulness (T), informativeness (I), and joint quality (T×I) on benchmarks. T : text modality; I : image modality. Red shading marks the top three methods per dataset.
Body System
MedMMV
CoT
T
I
T × I
T
I
T × I
Skeletal
4.42 ± 0.34
4.13 ± 0.40
73.26 ± 12.87
4.11 ± 0.17
4.47 ± 0.38
73.39 ± 6.12
Reproductive
4.75 ± 0.23
4.00 ± 0.01
76.17 ± 3.60
3.60 ± 0.40
4.66 ± 0.27
67.23 ± 11.38
Cardiovascular
4.28 ± 0.46
3.97 ± 0.26
68.28 ± 11.95
3.15 ± 0.58
4.02 ± 0.33
51.12 ± 13.13
Lymphatic
4.23 ± 0.45
4.18 ± 0.39
71.28 ± 13.83
3.87 ± 0.60
3.98 ± 0.67
62.69 ± 20.83
Nervous
4.27 ± 0.55
3.68 ± 0.30
63.31 ± 13.41
3.18 ± 0.42
4.23 ± 0.63
54.59 ± 15.35
Table 4: Physician evaluation on the MedXpert-MM dataset by body system. Each body system includes 20 samples; the per-doctor breakdown is in Appendix J.3 .
Table 5: Ablation study on MedXpert-MM and MedXpert-Text. Full ablations on six benchmarks are in Table 9 .
Method
Filtered
Pairwise
On-dist.
Repair
Acc
T
I
T × I
Imitation
Base (no training)
✗
✗
✗
✗
10.50
1.68
2.53
18.22
Trace-Imitation (raw CoT)
✗
✗
✗
✗
20.00
1.93
2.79
23.41
Trace-Imitation (Teacher-Trajectory)
✗
✗
✗
partial
24.00
2.18
2.81
26.18
Preference
RS-Imitation (Self-Consistency)
weak
✗
✗
✗
23.00
2.08
2.90
26.00
Table 6: Signal ladder: from trace imitation to closed-loop feedback transfer on MedXpert-MM (Qwen2.5-VL-7B student, MedMMV GPT-5 teacher). Each row adds a richer pipeline signal. Filtered : CU-based trajectory filtering; Pairwise : chosen/rejected preference objective; On-dist. : student generates rollouts; Repair : repair-aware scoring.
Figure 3: Pipeline feedback signals are calibrated for reasoning quality. (a) CU score increases with teacher accuracy, indicating alignment with true trajectory quality. (b) Higher-CU trajectories exhibit fewer hallucination issues, demonstrating that CU captures factual grounding. (c) Teacher reranking scores strongly predict student success, establishing teacher feedback as an effective reward signal for on-policy learning.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Instability breeds hallucination. Panels 1–4: Prior instability strongly predicts hallucination risk. (1) Higher entropy at step i−1 increases the probability of high CMHR at step i . (2) Majority-path churn accumulates globally and correlates with higher mean CMHR. (3) A single majority-path switch produces an immediate +13.3% increase in hallucination probability. (4) Timelines show that high-churn questions propagate instability across reasoning steps, while low-churn questions remain stable.
Figure 5: Instability correlates with elevated hallucination risk. Left (Panels 1–2) : Higher prior-step entropy is associated with increased probability of high CMHR, and majority-path churn correlates with higher mean CMHR across questions. Right : A multi-rollout case study illustrates how run-level instability (measured by RGM) coincides with divergent evidence interpretation, with unstable forks leading to inconsistent conclusions while stable rollouts converge.
Figure 6: Assumed mediation structure with controls X .
Dataset
Task
Modality
Original size
Subset size
MedXpertQA-MM
VQA
Image + Text
10,868
238
MedFrameQA
VQA
Image + Text
2,851
200
PathVQA
VQA
Image + Text
32,799
200
MedXpertQA-Text
QA
Text
26,675
200
MedMCQA
QA
Text
193,155
200
MedQA
QA
Text
12,723
200
Appendix
Table 7: Statistics of the six medical QA/VQA benchmarks used in our study. Each dataset includes the original size and the representative subset size used for evaluation, chosen to balance coverage and computational feasibility.
Question
A 36-year-old male patient presents with fever, pain, and wound drainage at the surgical site 4 months following an Achilles tendon repair. Figure A shows the clinical presentation. Laboratory testing reveals an elevated ESR of 29 mm/hr (normal range: 0-22 mm/hr). During the planned debridement and irrigation procedure, it is discovered that the Achilles tendon has re-ruptured with a 5 cm gap. Given these findings, what is the most appropriate next step in surgical management? Options: A: Primary repair of the re-ruptured Achilles tendon B: Repair of the re-ruptured Achilles tendon with a turndown procedure C: Repair of defect with flexor hallucis longus tendon transfer D: Debridement of necrotic and infected tendon tissue, with no attempt at reconstruction E: V-Y plasty of the re-ruptured Achilles tendon
Appendix
Table 8: Comparing the Outputs of Direct CoT and MedMMV on GPT-oss-120B.
Figure 7: Case study example of MedMMV with intermediate outputs.
Table 9: Ablation study on all six datasets. Note: Acc: Accuracy (%); T: Truthfulness (1–5); I: Informativeness (1–5); T × I: normalized product in [0,100], computed as (T×I)/25×100 . ImageDoctor ablations are not applicable (–) for text-only tasks. Bold indicates full MedMMV with best performance. Each dataset samples 200 cases.
Table 11: Hallucination metrics on MedXpertQA MM across medical categories.
Metric
Case-Level ρ
p-value
TRUE
0.86
<0.001
INFO
0.82
<0.001
TRUE × INFO
0.89
<0.001
Appendix
Table 12: Spearman correlation ( ρ ) between human and LLM evaluations across metrics.
Body System
ICC(T)
ICC(I)
Δ (T × I) Human
Δ (T × I) LLM
Skeletal
0.92
0.88
9.87
9.36
Reproductive
0.90
0.85
8.94
9.14
Cardiovascular
0.89
0.83
9.78
10.20
Lymphatic
0.91
0.87
8.74
9.03
Nervous
0.86
0.80
9.56
8.92
Digestive
0.88
0.82
10.24
10.48
Appendix
Table 13: Inter-rater reliability (ICC) and improvement consistency across body systems. ICC(T) and ICC(I): Intraclass Correlation Coefficients for Truthfulness and Informativeness among three physicians per system. Δ (T × I): improvement of MedMMV over CoT, measured by human raters and LLM judges respectively.
Analysis
Setting
Result
Key Observation
Human reliability
ICC(T) / ICC(I)
0.89 / 0.85
Good-to-excellent agreement
Evidence robustness
Noise 0–80%
AURC = 64.1 (Acc)
Smooth decay, not brittle
CU weight sensitivity
wev , wcoh variation
≤ 1% Acc change
Stable across configurations
Efficiency
Per-query latency
40–62s
Practical for non-emergent use
Input variation
Resolution / Context
Stable / Sensitive
Context length is critical
Appendix
Table 14: Summary of reliability and robustness analyses on MedXpert-MM (GPT-5).
Metric
0%
20%
40%
60%
80%
Δ (80–0)
AURC ( ↑ )
Accuracy
73.58
73.59
72.98
54.26
37.71
–35.87
64.12
T × I
82.89
81.99
78.72
62.53
52.22
–30.67
72.70
Appendix
Table 15: Robustness to evidence graph noise on MedXpert-MM (GPT-5). Mixed structural and semantic perturbations are injected at increasing levels. AURC: area under the robustness curve (higher is better).
Statistic
Sevidence
Scoherence
Prepair
CU
Mean
0.78
0.72
0.31
1.19
Std. Dev.
0.12
0.15
0.09
0.18
Min
0.50
0.40
0.10
0.80
Max
0.95
0.95
0.50
1.50
Appendix
Table 16: Distribution of scores for GPT-5 on the CU evaluation metrics.
Module
Parameter
Value
Acc
T
I
T × I
CU Weights
wevidence
1
73.58
4.26
4.73
82.89
3
73.16
4.31
4.70
82.10
5
74.00
4.29
4.73
82.69
wcoherence
1
73.58
4.26
4.73
82.89
3
74.26
4.19
4.82
82.81
5
73.28
4.23
4.79
82.93
Appendix
Table 17: Sensitivity analysis of CU scorer weights and key hyperparameters on MedXpert-MM (GPT-5). Boldface marks the default configuration used throughout the paper.
Stage
Input Tok.
Output Tok.
GPT-5 (s)
Qwen-7B (s)
Stage 1 – Short rollouts
≈ 1,011
≈ 5,058
8.4
4.5
Stage 2 – Reasoning expansion
≈ 5,058
≈ 14,586
24.3
20.2
Stage 3 – Decision aggregation
≈ 1,710
≈ 4,287
21.4
11.7
Hallucination detector
≈ 1,510
≈ 1,553
7.8
4.1
Total
≈ 9,289
≈ 25,484
61.9
40.5
Appendix
Table 18: Per-stage token usage and inference latency. Times are averaged over MedXpert-MM queries accessed via official APIs.
Category
Setting
Acc
T
I
T × I
Image Resolution
20 dpi
73.11
4.27
4.68
82.01
50 dpi
74.79
4.30
4.71
82.79
100 dpi
73.58
4.26
4.73
82.89
Context Length
500 tokens
27.73
2.68
3.53
48.22
1,000 tokens
60.08
3.01
4.09
67.92
2,000 tokens
71.65
3.97
4.70
79.23
Appendix
Table 19: Robustness to input variations on MedXpert-MM (GPT-5). Performance under different image resolutions and context length constraints.
Group
CU Score ↑
Fact Check ↑
Halluc. Issues ↓
Teacher Acc ↑
Rejected
0.65
0.42
68.3
61.2%
Accepted
1.12
0.71
38.7
77.6%
Appendix
Table 20: Quality comparison between accepted and rejected teacher trajectories during CU-based rejection sampling on MedXpert-MM (GPT-5 teacher, 134 training samples).
Data Size
Accuracy (%)
Evid. Support
Raw CoT
MedMMV-filtered
Raw CoT
MedMMV-filtered
34 (25%)
20.0
20.0
0.30
0.37
67 (50%)
21.0
21.0
0.31
0.38
100 (75%)
22.0
21.0
0.32
0.39
134 (100%)
23.0
21.0
0.32
0.39
Appendix
Table 21: Effect of training data size on SFT performance (MedXpert-MM, Qwen2.5-VL-7B). MedMMV-filtered supervision is compared with raw CoT traces at matched data scales.
Table 23: Detailed experimental settings (Part I: shared configuration and offline training signals).
Figure 8: Why filtered and on-policy feedback work. (a) Selected teacher trajectories have higher CU scores than rejected ones, confirming that CU-based filtering retains higher-quality reasoning paths. (b) Selected trajectories also exhibit fewer hallucination issues, indicating that the filtering signal is aligned with factual quality. (c) Teacher reranking scores on student rollouts show clear separation between low- and high-quality trajectories, providing a meaningful feedback signal for on-policy learning.
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.
Hao Fang, Jinyu Li, Jiawei Kong +4
Tsinghua Shenzhen International Graduate School, Tsinghua University · Harbin Institute of Technology, Shenzhen
Reinforcement Learning with Verifiable Rewards (RLVR) improves multimodal reasoning by rewarding verifiable final answers. Yet answer-correct trajectories may still rely on incomplete derivations, weak evidence, or statements that contradict their conclusions. This gap between answer correctness and reasoning validity, which we call reasoning-answer inconsistency, motivates trajectory supervision in multimodal RL. We compare two main approaches: reward models (RMs), and Generative Rewards (GRs). RMs are efficient and help early in training, but their gains weaken as the policy distribution shifts; GRs improve performance, but may give unstable rewards and computationally expensive. We therefore propose Groupwise Ranking Reward, which ranks verifier-passed trajectories for the same prompt in one pass and redistributes reward accordingly. Groupwise comparison better separates stronger and weaker correct trajectories with lower judge overhead than GRs. Experiments show that RLVR aggravates reasoning-answer inconsistency, while trajectory supervision alleviates it. Groupwise Ranking Reward performs best overall, improving reliability-conditioned accuracy from 47.4% to 54.7% over RLVR.
Multimodal large language models (MLLMs) show strong promise for clinical VQA and radiology report generation, yet inference-time hallucinations still undermine trustworthy use: models can produce fluent conclusions that conflict with imaging evidence. Existing mitigation strategies typically rely on additional training, external retrieval/knowledge bases, or multi-stage post-hoc verification, which increases cost and pipeline complexity and often generalizes poorly across models and tasks.To address this, we propose a holistic, training-free evidence-injection framework that systematically mitigates hallucinations through dual-side evidence injection. By leveraging ROI priors acquired using MedSAM in our implementation, we recalibrate the visual perception trajectory via ROI-guided activation modulation while anchoring the textual reasoning trajectory by mapping anatomical coordinates into discrete semantic tokens as verifiable external memory. Then we introduce a task-aware dynamic router to select modality-specific interventions based on task semantics, balancing perceptual grounding and linguistic fluency. We conduct systematic evaluations on 2 tasks and 5 datasets using LLaVA-1.5-7B, LLaVA-Med-1.5-7B, Qwen3-VL-8B/32B, and InternVL-3.5-8B/38B. Controlled ablations and visualizations further validate the framework, which consistently outperforms baselines across medical benchmarks, improving close-ended accuracy by up to ∼6%↑ and reducing open-ended hallucinations by ∼35%↓. The code has been made available on GitHub: blue{https://github.com/Henry991115/SPRG}.
Rui Hao, Qiankun Li, Junyuan Mao +4
Huazhong University of Science and Technology, Wuhan, China · Imperial Global Singapore, Imperial College London, London, United Kingdom · Nanyang Technological University, Singapore, Singapore +2