CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework
Authors: Yuexi Du, Jinglu Wang, Shujie Liu, Nicha C. Dvornek, Yan Lu
Organizations: Microsoft Research Asia · Department of Biomedical Engineering, Yale University · Department of Radiology & Biomedical Imaging, Yale University
Large visual language models (VLMs) have shown strong multi-modal medical reasoning ability, but most operate as end-to-end black boxes, diverging from clinicians' evidence-based, staged workflows and hindering clinical accountability. Complementarily, expert visual grounding models can accurately localize regions of interest (ROIs), providing explicit, reliable evidence that improves both reasoning accuracy and trust. In this paper, we introduce CARE, advancing Clinical Accountability in multi-modal medical Reasoning with an Evidence-grounded agentic framework. Unlike existing approaches that couple grounding and reasoning within a single generalist model, CARE decomposes the task into coordinated sub-modules to reduce shortcut learning and hallucination: a compact VLM proposes relevant medical entities; an expert entity-referring segmentation model produces pixel-level ROI evidence; and a grounded VLM reasons over the full image augmented by ROI hints. The VLMs are optimized with reinforcement learning with verifiable rewards to align answers with supporting evidence. Furthermore, a VLM coordinator plans tool invocation and reviews evidence-answer consistency, providing agentic control and final verification. Evaluated on standard medical VQA benchmarks, our CARE-Flow (coordinator-free) improves average accuracy by 10.9% over the same size (10B) state-of-the-art (SOTA). With dynamic planning and answer review, our CARE-Coord yields a further gain, outperforming the heavily pre-trained SOTA by 5.2%. Our experiments demonstrate that an agentic framework that emulates clinical workflows, incorporating decoupled specialized models and explicit evidence, yields more accurate and accountable medical AI. Project page: https://xypb.github.io/CARE-Project-Page/
Figures & tables
Figure 1: VLMs for medical reasoning. (a) Single-shot VLMs often miss local evidence. (b) Grounding VLMs do not explicitly utilize ROI in reasoning. (c) Generalist visual reasoning VLMs fail with incorrect initial focus. (d) Our agentic CARE-Coord performs grounded evidence-based reasoning and expert discussion, improving accountability. (e) Comparison of average medical VQA accuracy vs. model size. Models with unknown size appear in the rightmost panel.
Figure 2: Method overview. The proposed CARE comprises a VLM coordinator and a set of task-specific expert models. The coordinator plans tool use and conducts answer review, invoking specialist models as needed. The expert set includes: (1) a question-conditioned entity-proposal VLM that identifies relevant anatomical structures/findings; (2) a referring segmentation model that localizes entities with pixel-level ROI evidence; and (3) an evidence-grounded VQA VLM that reasons over the image augmented with selected visual evidence (zoom-in, mask, or global indicator).
Model
OMVQA-3k
VQA-RAD
SLAKE
VQA-Med-2019
Overall
Proprietary
GPT-4o ( Hurst et al., 2024 )
64.07
58.54
63.55
59.60
61.44
GPT-5 ( OpenAI, 2025 )
74.73
63.19
67.75
62.20
66.97
Open-source
Llama-3.2-11B-Vision ( Meta, 2024 )
43.10
53.22
63.17
57.40
54.22
Qwen2.5-VL-7B ( Bai et al., 2025 )
61.40
54.10
59.73
50.60
56.46
Table 1: Quantitative results on medical VQA benchmarks. We report medical VQA accuracy (%) on four standard benchmarks: OMVQA-3k ( Hu et al., 2024 ) , VQA-RAD ( Lau et al., 2018 ) , SLAKE ( Liu et al., 2021 ) and VQA-Med-2019 ( Ben Abacha et al., 2019 ) . Open-ended questions are scored by GPT-4o against ground-truth answers. Our segmentation model is smaller than 1B. We highlight medical expert VLMs in gray and ours in green .
Training Visual Clue
Coordinator
Datasets
ID
OOD
Overall
Mask
Zoom
Global
Planning
Review
OMVQA
VQA-RAD
SLAKE
Avg.
VQA-Med-2019
94.5
60.5
78.8
77.9
56.0
72.4 ~{}_{{\color[rgb]{0.5,0.5,0.5}+0.0}}
✓
95.4
61.6
81.8
79.6
54.0
73.2 ~{}_{{\color[rgb]{1,0,0}+0.8}}
✓
95.1
61.2
82.3
79.5
56.8
73.8 ~{}_{{\color[rgb]{1,0,0}+1.4}}
✓
✓
95.6
62.0
83.1
80.2
55.6
74.1 ~{}_{{\color[rgb]{1,0,0}+1.7}}
Table 2: Ablation on grounded VQA. We ablate the 8B EG-VQA components during training, varying training visual evidence and coordinator configurations. Only one type of visual evidence is used for inference. CARE-Flow and CARE-Coord are highlighted in blue and green , respectively.
Table 5Table 6Table 7
Method
VQA-RAD
SLAKE
Open Recall
Closed Accuracy
Open Recall
Closed Accuracy
FAVP - Vicuna ( Yu et al., 2025b )
71.90
88.20
87.20
88.10
BioMed-VITAL-13B ( Cui et al., 2024 )
69.72
84.86
91.69
90.70
PMC-CLIP ( Lin et al., 2023 )
-
84.00
-
88.00
Ours (CARE-Coord-B)
66.27
78.88
87.34
89.19
Table 9: Comparison with More Baselines. We compare with additional baselines using their reported metrics (Recall for open-ended, Accuracy for closed-ended). FAVP and BioMed-VITAL use larger pre-trained data and fine-tune per dataset. PMC-CLIP is a classification-based method incompatible with open-ended generation. Our method is highlighted in green .
Figure 3: Case Study. We present the complete reasoning trace for a CT disease identification question. Key information from the coordinator is highlighted in blue , model reasoning in green , and each model’s final answer in yellow .
Table 10: Medical Training Data for each Model. We report the medical-specific training data for each baseline for fair comparison. We highlight the overlapped training data with ours in bold . We highlight results of medical expert VLMs using gray , and our model using green
Figure 4: Prompt for Data Synthesis. We present the prompt used for the GPT-4o model to synthesize training data for the entity proposal model. We ask the model to generate questions based on the given meta-information of the provided image. The question is related to the medical entity/ies in the metadata.
Figure 5: Example Metadata for Data Synthesis. We present the metadata used for the GPT-4o model to synthesize training data for the entity proposal model. It includes the information about the original image, medical entities labeled from the dataset, and other related information, like the position of each mask.
Figure 6: Reward Curve During Training We provide the full reward curve of the entity proposal model during training. (a) is the overall reward, and (b) is the individual similarity reward.
Figure 7: Reward Curve During Training We provide the full reward curve of the G-VQA model during training. (a) is the overall reward, and (b) is the individual accuracy reward.
τC
Overall Avg. Acc.
InternVL3-2B
InternVL3-8B
0%
70.60
74.16
30%
70.82
74.66
50%
71.02
74.84
70%
71.14
74.92
100%
69.72
72.48
Appendix
Table 11: Confidence Threshold Ablation. We ablate different mask dropping confidence thresholds during training. We use CARE-Flow model here to isolate the influence of the coordinator and only focus on different mask confidence thresholds. Our choice of τC is highlighted in green .
Model
Avg. Tool Invocations
Inference Time Per-input (seconds)
Coordinator
Entity Prop
Segmentation
GVQA
Overall
Fixed Visual Clue
3.00
-
1.31
0.06
2.03
3.44
Majority Vote (CARE-Flow)
5.00
-
1.27
0.06
5.81
7.22
InternVL3-38B
3.09
11.57
1.06
0.09
1.97
18.63
GPT-4o
2.82
17.44
1.24
0.07
2.08
21.08
GPT-5 (CARE-Coord)
2.50
39.32
1.09
0.12
2.28
43.51
Appendix
Table 12: Inference Speed Evaluation We compute the time cost for each component in our CAREmodel during inference time on OmniMedVQA ( Hu et al., 2024 ) dataset. We report the time in seconds used to finish one single VQA request.
Model
OMVQA-3k
VQA-RAD
SLAKE
VQA-Med-2019
Overall
Proprietary
GPT-4o ( Hurst et al., 2024 )
64.07
58.54
63.55
59.60
61.44
GPT-5-mini ( OpenAI, 2025 )
68.57
59.87
65.94
61.60
63.99
GPT-5 ( OpenAI, 2025 )
74.73
63.19
67.75
62.20
66.97
Open-source
Llama-3.2-11B-Vision ( Meta, 2024 )
43.10
53.22
63.17
57.40
54.22
Appendix
Table 13: Full Benchmarking Results on Medical VQA Datasets. We present the full benchmarking results on 4 Medical VQA datasets ( Hu et al., 2024 ; Liu et al., 2021 ; Ben Abacha et al., 2019 ; Lau et al., 2018 ) and report their accuracy (%). The open-ended questions are evaluated using GPT-4o ( Hurst et al., 2024 ) against ground-truth. We use the Instruct fine-tuned model whenever available. We use a 3k subset of OmniMedVQA (OMVQA) Hu et al. (2024) dataset for benchmark. We highlight results of medical expert VLMs using gray , and our model using green .
Model
OMVQA-3k
VQA-RAD
SLAKE
VQA-Med-2019
Overall
InternVL3-2B (zero-shot)
75.43
55.65
63.07
54.00
62.04
InternVL3-2B-Finetuned
87.97
57.43
69.56
51.00
66.49
CARE-Flow-S
94.53
56.32
78.44
53.60
70.72
InternVL3-8B (zero-shot)
75.97
61.86
66.13
57.40
65.34
InternVL3-8B-Finetuned
91.13
61.86
76.53
53.80
70.83
CARE-Flow-B
96.17
63.64
83.21
56.60
74.91
Appendix
Table 14: Comparison with Fine-tuned Baseline. We report medical VQA accuracy (%) on four standard benchmarks: three in-domain OMVQA-3k ( Hu et al., 2024 ) , VQA-RAD ( Lau et al., 2018 ) , SLAKE ( Liu et al., 2021 ) and one out-of-domain VQA-Med-2019 ( Ben Abacha et al., 2019 ) . We compare with the InternVL3 ( Zhu et al., 2025a ) baseline fine-tuned with the same training data to isolate the influence of different training data. Our results are highlighted in blue .
Model
OMVQA-3k
VQA-RAD
SLAKE
VQA-Med-2019
Overall
σ=25
96.70
66.29
83.77
56.40
75.79
σ=50
96.74
67.40
83.87
56.60
76.15
σ=75
97.47
67.85
83.68
56.80
76.45
CARE-Coord-B
97.97
68.29
83.11
60.80
77.54
Appendix
Table 15: Conservative Coordinator Review Strategy. We report medical VQA accuracy (%) on four standard benchmarks ( Hu et al., 2024 ; Lau et al., 2018 ; Liu et al., 2021 ; Ben Abacha et al., 2019 ) to compare our coordinator strategy against a conservative strategy that votes for EG-VLM’s answer when its confidence is high. We report the performance with different confidence threshold σ . Our results are highlighted in green .
LLM Judger
VQA-RAD
SLAKE
VQA-Med-2019
Overall
GPT-4o
68.29
83.11
60.80
70.73
GPT-4o-mini
68.51
84.35
60.60
71.15
InternVL3-38B-Instruct
67.41
82.44
59.00
69.62
InternVL3-78B-Instruct
67.41
83.30
60.20
70.30
Avg.
67.91
83.30
60.15
70.45
STD
± 0.58
± 0.79
± 0.81
± 0.66
Appendix
Table 16: Different LLM-as-Judge. We report the performance on the datasets with open-ended questions Lau et al. (2018) ; Liu et al. (2021) ; Ben Abacha et al. (2019) to demonstrate the variance of different LLM-as-judge. We experimented with both proprietary ( Hurst et al., 2024 ) and open-source ( Zhu et al., 2025a ) LLMs of different sizes. We report the averaged performance across different LLMs and the corresponding standard deviation.
Method
Entity Accuracy
Mask Dice
Overall MedVQA
SFT
76.70
72.69
74.26
SFT + RFT
80.92
74.03
74.44
RFT (CARE-Flow-B)
85.28
73.48
74.91
Appendix
Table 17: Entity Proposal Model Fine-tuning Strategy Ablation. We ablate different fine-tuning strategies for the Entity proposal VLM on the entity proposal, referring entity segmentation, and medical VQA tasks. Our choice of training strategy is highlighted in blue .
Coordinator
Human Evaluation Pass Rate (%)
GPT-4o
73.94
GPT-5 (CARE-Coord-B)
82.14
Appendix
Table 18: Reasoning Trace Human Evaluation. We conduct a human evaluation that evaluates the quality of the reasoning trace of our model on a subset of samples. We report the pass rate of human evaluation. We compare our CARE-Coord-B against our model with the GPT-4o coordinator. Our method is highlighted in green .
Figure 8: System Prompt for Entity Proposal Model. We present the system prompt for our entity proposal model, where we instruct the model to name the most relevant medical entity related to the user question. We note that even if we asked the model to also generate the size and position of the entity, we only use the entity name for the downstream tasks.
Figure 9: System Prompt for G-VQA Model. We present the system prompt for our grounded VQA model, where we provide information about different types of visual clues and guide the model to focus differently when given these visual clues.
Figure 10: System Prompt for Coordinator Model. We present the system prompt for our coordinator model. We introduce the overview of each tool and define the general workflow here. The section highlighted in blue defines the coordination and planning behavior, and the section highlighted in green defines the iterative answer review process during inference.
Figure 11: Example of Synthetic Data for Entity Proposal Task. An example of synthetic data for the entity proposal task, we wrap the ground truth entity name in the JSON object.
Figure 12: Example of Synthetic Data for Entity Proposal Task. An example of synthetic data for the entity proposal task, we wrap the ground truth entity name in the JSON object.
Figure 13: Example of Synthetic Data for Entity Proposal Task. An example of synthetic data for the entity proposal task, we wrap the ground truth entity name in the JSON object.
Figure 14: More Case Study. We provide a full reasoning process of a user’s question about a brain MRI image here. We highlight the key information from the coordinator in blue and key information from the tool in green.
Figure 15: More Case Study. We provide a full reasoning process of a user’s question about a Chest X-ray image here. We highlight the key information from the coordinator in blue and key information from the tool in green.
Figure 16: More Case Study. We provide a full reasoning process of a user’s question about a body CT image here. We highlight the key information from the coordinator in blue and key information from the tool in green. This example skipped the entity proposal and segmentation, as the question is about global information of the image.
Figure 17: More Case Study. We provide a full reasoning process of a user’s question about a digital image here. We highlight the key information from the coordinator in blue and key information from the tool in green.
Figure 18: More Case Study. We provide a full reasoning process of a user’s question about a Chest X-ray here. We highlight the key information from the coordinator in blue and key information from the tool in green.
Figure 19: Case Study. We present the complete reasoning trace for a CT question. Key information from the coordinator is highlighted in blue , model reasoning in green , and each model’s final answer in yellow . Notably, during review, the coordinator corrects an initial error from the entity proposal.
Figure 20: Failed Case Study. We provide a full reasoning process of a user’s question about a CT image with a failed result here. We highlight the reason for failure during the reasoning chain, where the G-VQA model gives a correct answer, but the coordinator model insists on a hallucinated answer. We use the red dashed arrows to illustrate how the hallucination propagates during the inference.
Figure 21: Function Definition for Entity Proposal Model We provide the full definition of the function schema of the entity proposal model used during coordination. We explain in detail the definition of each parameter here.
Figure 22: Function Definition for Segmentation Model We provide the full definition of the function schema of the referring segmentation model, which accepts two inputs and outputs the segmentation mask along with its confidence.
Figure 23: Function Definition for G-VQA Model We provide the full definition of the function schema of the G-VQA model. Note that we only set the input image, question, and clue type to be required parameters, as when using a global clue type, there is no need to use other parameters about the mask.
Figure 24: User Interface of Human Evaluation. We provide the full user interface of our human evaluation.
Reinforcement Fine-Tuning (RFT) has enabled medical Multimodal Large Language Models (MLLMs) to produce Chain-of-Thought (CoT) reasoning for visual question answering, yet these models suffer from confidence miscalibration---a systematic gap between expressed certainty and actual diagnostic accuracy that undermines clinical trust. We propose CARE, a Confidence-Aware medical REasoning framework that jointly optimizes accuracy and calibration through a dual-stage pipeline. First, a scalable Medical-CoT synthesis provides structured cold-start data for Supervised Fine-Tuning. Second, Group Relative Policy Optimization (GRPO) with a novel Confidence-Aware Reward (CAR) mechanism ties the model's confidence to diagnostic correctness within the reward signal. Across three Medical VQA benchmarks, CARE achieves the highest diagnostic accuracy while obtaining the lowest Expected Calibration Error and Hallucination Rate, establishing a foundation for trustworthy clinical decision support. Our code is available at https://github.com/anotherbricki/CARE.
Yuetian Du, Yucheng Wang, Zhenyuan Chen +9
Zhejiang University · Ant Group · University of Michigan +1
Medical vision-language models (VLMs) show strong performance on radiology tasks but often produce fluent yet weakly grounded conclusions due to over-reliance on a dominant modality. We introduce a context-aligned reasoning framework that enforces agreement across heterogeneous clinical evidence before generating diagnostic conclusions. The proposed approach augments a frozen VLM with structured contextual signals derived from radiomic statistics, explainability activations, and vocabulary-grounded semantic cues. Instead of producing free-form responses, the model generates structured outputs containing supporting evidence, uncertainty estimates, limitations, and safety notes. We observe that auxiliary signals alone provide limited benefit; performance gains emerge only when these signals are integrated through contextual verification. Experiments on chest X-ray datasets demonstrate that context alignment improves discriminative performance (AUC 0.918 to 0.925) while maintaining calibrated uncertainty. The framework also substantially reduces hallucinated keywords (1.14 to 0.25) and produces more concise reasoning explanations (19.4 to 15.3 words) without increasing model confidence (0.70 to 0.68). Cross-dataset evaluation on CheXpert further reveals that modality informativeness significantly influences reasoning behavior. These results suggest that enforcing multi-evidence agreement improves both reliability and trustworthiness in medical multimodal reasoning, while preserving the underlying model architecture.
Sumra Khan, Sagar Chhabriya, Aizan Zafar +5
Department of Computer Science, Salim Habib University, Karachi, Pakistan · Computer Science, Institute of Business Administration Sukkur, Pakistan · Center for Research in Computer Vision, University of Central Florida, USA +2
Vision-language models have demonstrated impressive capabilities in general medical visual question answering, yet due to limited interpretability, it remains unclear whether their predictions reflect evidence-grounded clinical reasoning or reliance on spurious priors. We introduce Med-R2 Bench, a hierarchical benchmark aligned with the clinical workflow to evaluate adversarial robustness with visual grounding. We design stepwise QA tasks to assess whether reasoning chains are strictly grounded in visual evidence across the four clinical stages, and employ adversarial perturbations to test robustness against misleading cues. Med-R2 comprises 42,432 images, 31 task categories, and 110,406 QA pairs. Evaluation across 14 VLMs reveals a sequential performance degradation along the four-stage clinical workflow. Adversarial experiments show that models rely heavily on correct prompts to guess answers. Even when provided with explicit visual cues, the models struggle to accurately align textual descriptions. Finally, we demonstrate stepwise fine-tuning using our hierarchical data significantly improves reasoning robustness, highlighting its potential to drive future improvements in evidence-based medical AI.
Wen Ma, Fucheng Niu, Zhiting Fan +3
Zhejiang University, Zhejiang, China · Guangdong Institute of Intelligence Science and Technology