Paper 2602.06570
Baichuan-M3: Modeling Clinical Inquiry for Reliable Medical Decision-Making
- Published
- Feb 2026
- Research lab
- Independent
- Citations
- 8
- GitHub
- 254 stars
01 In brief
Summary
Baichuan-M3 is a medical-enhanced large language model designed to shift from passive question-answering to active, clinical-grade decision support.
It addresses limitations in open-ended consultations by modeling the systematic workflow of a physician, with key capabilities including proactive information acquisition, long-horizon reasoning, and adaptive hallucination suppression.
The model uses a three-stage training pipeline: Task-Specific Reinforcement Learning (TaskRL), Offline Policy Distillation, and Multi-Teacher Online Policy Distillation (MOPD).
For long-horizon consultations, it introduces Segmented Pipeline Reinforcement Learning and the SPAR algorithm for step-wise credit assignment.
A Fact-Aware Verification Pipeline with a two-level caching system reduces hallucination rates.
Baichuan-M3 achieves state-of-the-art results on HealthBench (44.4 on HealthBench-Hard), HealthBench-Hallu, and the new ScanBench benchmark, outperforming GPT-5.2 in clinical inquiry, advisory, and safety.
Inference optimizations include Gated Eagle-3 speculative decoding and INT4 quantization.
The model is publicly available on Hugging Face.
02 From the paper
Abstract
We introduce Baichuan-M3, a medical-enhanced large language model engineered to shift the paradigm from passive question-answering to active, clinical-grade decision support. Addressing the limitations of existing systems in open-ended consultations, Baichuan-M3 utilizes a specialized training pipeline to model the systematic workflow of a physician. Key capabilities include: (i) proactive information acquisition to resolve ambiguity; (ii) long-horizon reasoning that unifies scattered evidence into coherent diagnoses; and (iii) adaptive hallucination suppression to ensure factual reliability. Empirical evaluations demonstrate that Baichuan-M3 achieves state-of-the-art results on HealthBench, the newly introduced HealthBench-Hallu and ScanBench, significantly outperforming GPT-5.2 in clinical inquiry, advisory and safety. The models are publicly available at https://huggingface.co/collections/baichuan-inc/baichuan-m3.