=== SUMMARY ===

Technical Summary: MIRA (Medical Image Reflection for Agentic Diagnosis)

1. Problem Statement

Recent advances in "thinking with images" have demonstrated that visual information can be actively leveraged during multimodal reasoning. However, indiscriminate tool use in medical agents introduces significant risks: noisy observations, the propagation of misleading evidence, and increased reasoning costs. In the high-stakes domain of medical diagnosis, merely acquiring additional observations is insufficient. A reliable agent must determine when external tools are necessary, verify whether acquired evidence supports the current hypothesis, and identify when weak or contradictory evidence necessitates corrective reasoning. Existing approaches often suffer from fixed visual observations (leading to missed lesions or hallucinations) or unregulated tool usage that distracts from the diagnostic process.

2. Methodology

The paper introduces MIRA (Medical Image Reflection for Agentic Diagnosis), a framework designed to convert static medical Visual Question Answering (VQA) into an iterative process of evidence seeking, tool-use verification, and self-correction. MIRA operates on four core principles: evidence-grounded visual interaction, tool-use reliability, failure-aware reflection, and continual principle evolution.

The technical implementation follows a two-stage training strategy:

Stage 1: MIRA-SFT (Cold Start)

This stage aims to teach the model tool use and reflection abilities through Supervised Fine-Tuning (SFT) on a carefully curated dataset.

• Tool Design: MIRA utilizes a lightweight suite of six tools: SEARCH (external knowledge), GROUNDING (bounding-box localization), POINT (fine-grained marking), ZOOM (region enlargement), ROTATE (orientation adjustment), and MEASURE (relative quantification).
• Golden Tool-Use Trajectory Synthesis: To bypass the cold-start exploration problem, the authors employ a tool-augmented Monte Carlo Tree Search (MCTS) data engine. Using frozen, high-capacity teacher models (Seed 1.8, Gemini 2.5 Pro, GPT-5.2), the system explores the diagnostic reasoning space. It generates diverse hypotheses, validates tool arguments via visual rendering, and scores intermediate states based on visual plausibility and semantic usefulness. This produces "golden" trajectories that are answer-correct, tool-valid, and evidence-aware.
• Reflection Thinking Pattern Synthesis: Beyond successful paths, the system synthesizes reflection data from two sources:
  1. Failure-driven Correction: For trajectories leading to incorrect answers, a teacher model identifies unstable observations or reasoning leaps and proposes corrective evidence-gathering actions.
  2. Textual Intention Reflection: For successful trajectories, the system verifies that each thought explicitly justifies the subsequent tool call or answer, rewriting unclear thoughts to ensure clear diagnostic goals.

Stage 2: MIRA-RL (Reinforcement Learning)

This stage optimizes the SFT policy using Group Relative Policy Optimization (GRPO) with on-policy rollouts.

• Composite Trajectory Reward: The reward function combines three components:
  1. Formatting Reward: Ensures adherence to output structures.
  2. Result Reward: Evaluates final answer correctness.
  3. Consistency Reward: A critical component that verifies whether the final answer is supported by the preceding visual/tool evidence and reasoning trajectory. This term is gated by the result reward (only active if the answer is correct).
• On-Policy Reflection Memory Optimization: To make failures reusable, the system implements a validation-gated memory update. Failed rollouts are distilled into generalizable diagnostic principles. These candidate principles are tested on a held-out validation set; they are only accepted if they improve the validation reward. This mechanism allows the model to learn from repeated failure patterns without overfitting to specific outliers.

3. Key Contributions

• MIRA Framework: A medical visual diagnostic framework that integrates autonomous evidence search with reflective verification, enabling agents to decide when to use tools and how to correct their reasoning.
• MCTS-Guided Data Engine: A novel data construction method that synthesizes high-quality, diverse diagnostic trajectories (both successful and failure-driven) using tree search and teacher models, addressing the lack of tool-use supervision in standard VQA datasets.
• Validation-Gated Reflection Memory: An online learning mechanism that distills repeated failures into reusable diagnostic principles, ensuring that memory updates are driven by performance improvements on held-out data rather than noise.
• Consistency-Aware Reward: A reward design that explicitly penalizes diagnostic jumps unsupported by visual evidence, encouraging evidence-grounded reasoning.

4. Experimental Results

Experiments were conducted across nine medical VQA benchmarks, including SLAKE, PMC-VQA, OmniMedVQA, MedLesionVQA, VQA-RAD, PathVQA, MedXpertQA-MM, and MMMU-Medical.

• Performance Gains: MIRA-VL-8B (based on the Qwen3-VL-8B backbone) consistently outperformed its base model, improving the average score from 57.29 to 64.73.
• Specific Improvements: The most significant gains were observed in fine-grained lesion understanding (MedLesionVQA: +14.5 points; MedLesionMCQ: +14.4 points) and expert-level reasoning tasks.
• Comparison: MIRA-VL-8B achieved competitive performance against larger open-source models (e.g., InternVL3-78B, Qwen3-VL-32B) and medical-specific models, narrowing the gap with larger closed-source systems like Gemini-2.5-Pro and GPT-5.2.
• Tool Use Efficiency: Analysis showed that MIRA increased "useful tool use" from 56.2% (in a basic tool-enabled baseline) to 73.8%, while reducing "harmful tool use" from 8.9% to 1.6%.
• Learning Dynamics: RL training resulted in shorter completion lengths and fewer tool turns, indicating the model learned to suppress redundant reasoning and only invoke tools when necessary.

5. Significance and Claims

The paper claims that MIRA demonstrates that evidence-grounded verification and corrective tool use can significantly improve the reliability of medical visual diagnostic agents. By moving beyond static image encoding to an iterative process of active evidence acquisition and self-correction, MIRA addresses the critical need for agents to verify their own hypotheses against visual reality.

The authors emphasize that their approach does not rely on heavy external vision modules but instead stimulates the model's intrinsic visual exploration capabilities through a lightweight tool suite and rigorous reflection mechanisms. The work suggests that for medical AI, the ability to re-examine visual evidence, correct premature conclusions, and adapt tool-use strategies is as important as the initial diagnosis.

Limitations: The authors modestly note that MIRA remains constrained by the perception and knowledge capabilities of its backbone model. If the base model fails to recognize subtle findings or lacks clinical knowledge, reflection can mitigate but not fully eliminate errors. Additionally, the reflection memory's generalization to rare diseases or unseen imaging styles requires further validation.