Research map / Vision-language and multimodal / Multimodal LLMs
Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Mentor-Intern Collaborative Search (MICS) to build medical multimodal CoT data (MMRP) and trains Chiron-o1 (InternVL3 base) with a three-stage curriculum SFT, reporting SOTA on medical VQA and reasoning benchmarks.
Problem
Medical MLLMs lack high-quality step-by-step chain-of-thought data, and existing RL-based medical reasoning models use fixed rewards that do not evaluate reasoning path quality, leading to superficial or hallucinated reasoning.
Approach
Mentor models (GPT-4o, Gemini 2.5 Pro, Qwen2.5-VL-72B) extend reasoning prefixes step by step; intern models (small open MLLMs at two temperatures) finish the solution, and the fraction of interns answering correctly serves as the MICS-Score to select the next step, yielding validated CoT data used in a three-stage curriculum of SFT.
Key results
- Chiron-o1-8B averages 69.2 across five medical VQA benchmarks, +5.7 over InternVL3-8B (63.5).
- Chiron-o1-8B reaches 58.4 accuracy and 49.4 MICS-Score on the MMRP reasoning test, versus 31.2 and 23.5 for MedVLM-R1.
- Chiron-o1-2B averages 65.5 on medical VQA, +8.1 over InternVL3-2B (57.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune InternVL3-8B as Chiron-o1-8B using the proposed curriculum and MMRP | InternVL3-8B · VQA-RAD | choice accuracy: 73.1 -> 76.8 | Helped |
| fine-tune InternVL3-8B as Chiron-o1-8B using the proposed curriculum and MMRP | InternVL3-8B · MMRP (Reasoning) | ACC: 31.2 -> 58.4 | Helped |
| add medical VQA datasets to MMRP training (ablation) | Chiron-o1-8B training · VQA-RAD | choice accuracy: +5% | Helped |
| add medical VQA datasets to MMRP training (ablation) | Chiron-o1-8B training · SLAKE | choice accuracy: +7.3% | Helped |
| apply quality control to MICS-searched reasoning paths (ablation) | Chiron-o1-8B training · MMMU (H&M) | choice accuracy: 49.3% to 54.6% (+5.3%) | Helped |
| apply quality control to MICS-searched reasoning paths (ablation) | Chiron-o1-8B training · MedXpertQA_MM | ACC: 22.6% to 24.2% (+1.6%) | Helped |
| use MICS instead of direct mentor-generated reasoning paths (ablation) | MMRP reasoning-data construction · four-benchmark average | average performance: 2.7% | Helped |
| add MMRP reasoning data (ablation) | Chiron-o1-8B training · four benchmarks | average performance: 3% | Helped |
| train with Stage 1 only instead of the full stage-wise curriculum (ablation) | Chiron-o1-8B · MMMU (H&M) | choice accuracy: 6.9% | Hurt |
| train with Stage 1+2 instead of the full stage-wise curriculum (ablation) | Chiron-o1-8B · MMMU (H&M) | choice accuracy: 6.2% | Hurt |
| train with Stage 3 only instead of the full stage-wise curriculum (ablation) | Chiron-o1-8B · MMMU (H&M) | choice accuracy: 3.5% | Hurt |
| omit Stage 3 from the training curriculum (ablation) | Chiron-o1-8B · medical reasoning benchmarks | average performance: 3.1% and 0.25% | Hurt |
Surprising findings
- Medical reasoning RL models (Med-R1, MedVLM-R1) lose basic VQA capability while focusing on reasoning.
Limitations
- MICS requires many costly API requests and compute to search reasoning paths.
- MMRP dataset is relatively small and needs further scaling.
- Chiron-o1-2B underperforms on MMMU (H&M) relative to InternVL3-2B; no human evaluation of reasoning quality.
Experimental setup
Datasets: MMRP Huatuo_PubMedVision_Alignment Huatuo_PubMedVision_InstructionTuning PMC-VQA VQA-RAD SLAKE PathVQA Radiopaedia cases
Benchmarks: VQA-RAD SLAKE PathVQA PMC-VQA MMMU (Health & Medicine) MedXpertQA_MM MMRP test set
Metrics: accuracy BERT-Score MICS-Score
Baselines: Med-R1 MedVLM-R1 ChestX-Reasoner HuatuoGPT-Vision-34B LLaVA-Med GMAI-VL Gemini-2.5-Pro GPT-4o InternVL3-8B InternVL3-2B MedReason HuatuoGPT-o1
Builds on: chain-of-thought prompting process reward models curriculum learning GRPO LoRA
Closest papers by meaning
- Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.92)
- M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.91)
- OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning (CVPR 2026 · similarity 0.91)
- Uni-Med: A Unified Medical Generalist Foundation Model For Multi-Task Learning Via Connector-MoE (NeurIPS 2024 · similarity 0.90)
- Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning (ICML 2025 · similarity 0.90)
- SMMILE: An expert-driven benchmark for multimodal medical in-context learning (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.