MIRA Paper

Medical visual reasoning with reflective tool use

MIRA: Medical Image Reflection
for Agentic Diagnosis

An agentic diagnostic framework that searches for evidence, verifies what its tools reveal, and corrects unreliable conclusions.

Shengzhi Wang1Jun Yang5Kai Wu1Xiaozhong Ji2,♠Yiwen Ye3Ziyang Chen3Mingliang Xiong1Wen Fang1Mingqing Liu1Mengyuan Xu1Miaoxuan Shan4Caiyan Liu1Bin He1Qingwen Liu1,†

  1. 1 Tongji University
  2. 2 Nanjing University
  3. 3 Northwestern Polytechnical University
  4. 4 People’s Public Security University of China
  5. 5 École Normale Supérieure — PSL

♠ Project Leader† Corresponding Author

Tool-level reflection

Look. Verify.
Reflect. Correct.

  1. 01
    Inspect globally.

    Begin with the complete image and the clinical question before selecting a tool.

  2. 02
    Question insufficient evidence.

    Check whether the current observation actually supports the working diagnosis.

  3. 03
    Refine with zoom.

    Acquire targeted visual evidence, then revise or retain the conclusion accordingly.

MIRA reflects on tool actions and evidence quality—not by displaying private reasoning, but by verifying whether acquired evidence is relevant, sufficient, and consistent before producing an answer.

Example of MIRA inspecting a medical image, recognizing insufficient evidence, and zooming to verify the finding
Targeted visual inspection turns an uncertain first observation into evidence-grounded correction.

Abstract

Evidence before conclusion.

Medical visual agents can use tools to inspect images and retrieve knowledge, yet indiscriminate tool calls may introduce noisy or misleading evidence. MIRA combines autonomous evidence search with reflective verification: it decides when tools are needed, evaluates whether their outputs support the current hypothesis, and corrects unreliable paths. Training pairs tool-augmented Monte Carlo Tree Search with jointly verified visual and semantic signals to generate supervised fine-tuning data, then evolves reusable reflective principles online during reinforcement learning. Across nine medical visual reasoning benchmarks, MIRA reaches a 64.73 average, improving its backbone by 7.44 score points while making tool use more useful and less harmful.

GroundingPointZoomRotateMeasureSearchReflect

64.73

Average score

+7.44 score points from 57.29 across nine benchmarks.

73.8%

Useful judgments

+17.6 percentage points from 56.2%. Useful combines correct and insufficient-but-relevant categories.

1.6%

Harmful judgments

Down 7.3 percentage points from 8.9%, indicating fewer misleading tool interactions.

+28.57

Fine-grained gain

Percentage-point improvement on VQA-RAD Understanding, Diagnosis, and Suggestion.

Qualitative cases

Correction becomes visible.

Three examples show how MIRA checks textual, visual, and retrieved evidence before finalizing a response.

Case 06Textual correction
Case 6 showing textual reflection and zoom-based correction for an abdominal aorta finding

After an uncertain global assessment, a targeted zoom reveals a focal high-density calcified focus along the abdominal aortic wall, enabling correction of the initial description.

Case 04Measurement
Case 4 comparing the measured long axis of a skin lesion with a fingernail reference

The measurement tool compares a 94-pixel lesion long axis with a 99-pixel fingernail reference, yielding a 0.95 ratio for a more grounded size judgment.

Case 01Zoom + search
Case 1 using zoom and external search to assess swelling around a finger joint

Zoom localizes erythema and swelling at an interphalangeal joint; external search then contextualizes the observed presentation before the response is finalized.

Method

Learn to search. Learn to reflect.

MIRA is developed in two complementary stages that improve both evidence acquisition and evidence verification.

Stage 01

MIRA-SFT

Tool-augmented MCTS explores diverse diagnostic hypotheses. During expansion, visual grounding and semantic consistency are jointly verified, producing higher-quality trajectories for supervised fine-tuning.

Tool-augmented Monte Carlo Tree Search pipeline with visual and semantic reflection for MIRA-SFT
Tool-augmented tree exploration and reflection-guided trajectory construction.
Stage 02

MIRA-RL

Failures encountered during reinforcement learning are distilled into candidate diagnostic principles. A principle enters reflection memory only when it improves held-out rollout reward under the same frozen policy, preventing unhelpful updates from accumulating.

Validation-gated online evolution of reflective principles in MIRA-RL
Validation-gated reflective-principle evolution during reinforcement learning.

Results

Stronger reasoning, safer tool use.

Performance gains are accompanied by a shift toward evidence that helps rather than distracts.

Across nine benchmarks

64.73 average

MIRA consistently improves over its Qwen3-VL-8B backbone across a broad medical visual reasoning evaluation.

Radar chart comparing MIRA with its backbone across medical visual reasoning benchmarks
Benchmark performance across complementary medical visual reasoning tasks.
Tool necessity

Right is useful. Down is safer.

MIRA reaches 73.8% useful judgments and only 1.6% harmful judgments. In this analysis, desired movement is rightward toward useful evidence and downward away from harmful evidence.

Tool-necessity plot showing MIRA moving right toward more useful and down toward fewer harmful tool judgments
Useful combines correct and insufficient-but-relevant categories.

Citation

Reference MIRA.

MIRA is available as arXiv:2608.10827 in Computer Vision and Pattern Recognition, with a cross-listing in Artificial Intelligence.

View on arXiv 
@misc{wang2026mira,
  title={MIRA: Medical Image Reflection for Agentic Diagnosis},
  author={Wang, Shengzhi and
          Yang, Jun and
          Wu, Kai and
          Ji, Xiaozhong and
          Ye, Yiwen and
          Chen, Ziyang and
          Xiong, Mingliang and
          Fang, Wen and
          Liu, Mingqing and
          Xu, Mengyuan and
          Shan, Miaoxuan and
          Liu, Caiyan and
          He, Bin and
          Liu, Qingwen},
  year={2026},
  eprint={2608.10827},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.10827}
}