Source-linked AI summary
Baichuan-M3: Modeling Clinical Inquiry for Reliable Medical Decision-Making
Baichuan-M3 Team, :, Chengfeng Dou, Fan Yang, Fei Li, Jiyuan Jia, Qiang Ju, Shuai Wang, Tianpeng Li, Xiangrong Zeng, Yijie Zhou, Hongda Zhang, Jinyang Tai, Linzhuang Sun, Peidong Guo, Yichuan Mo, Xiaochuan Wang, Hengfu Cui, Zhishou Zhang
TL;DR
Existing medical LLMs face limitations in open-ended consultations, motivating systems that can acquire missing information while sustaining reliable, long-horizon reasoning. Baichuan-M3 uses workflow-aligned training and multi-teacher distillation to unify clinical inquiry with decision support. It achieves strong performance across factual and process-oriented benchmarks, although its current scope remains episodic and text-based.
Problem
Existing systems have limitations in open-ended clinical consultations, where reliable decision support requires more than passive question answering.
Method
Baichuan-M3 uses workflow-aligned training, including multi-teacher distillation, to support proactive information acquisition and coherent multi-step clinical reasoning.
Results
Baichuan-M3 achieves strong performance across factual and process-oriented benchmarks, including HealthBench, HealthBench-Hallu, and OSCE-style ScanBench.
Takeaways & Limitations
The results support workflow-aligned optimization as an effective approach for combining clinical inquiry with reliable medical decision-making.
Takeaways & Limitations
Baichuan-M3 is limited to episodic, text-based clinical scenarios and does not fully capture longitudinal disease management, multimodal signals, or ultra-long-horizon reasoning.
Abstract
from arXiv · showhide
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.
1 Introduction
Existing medical LLMs are evaluated mainly through disconnected factuality benchmarks or consultation simulations, leaving a gap between active inquiry and deep, evidence-grounded reasoning. Baichuan-M3 addresses this gap by modeling clinical workflows through proactive information acquisition, coherent reasoning, and adaptive hallucination suppression.
- Motivation: Existing benchmarks often separate single-turn factuality from multi-turn consultation performance, despite harder cases requiring complex clinical reasoning.HealthBench and Med-HALT measure factual consistency and hallucinations, while other evaluations focus on process-oriented consultation simulations.
- Motivation: Current systems struggle to unify conversational interaction with clinical reasoning as one coherent capability.Knowledge-centric models may fail to elicit missing evidence, whereas interaction-focused models may prioritize fluency over principled differential reasoning.
- Challenges: Long-horizon clinical interactions create bottlenecks in multi-task fusion, reinforcement-learning credit assignment, and reward design.Terminal supervision obscures which conversational turns matter, while reward saturation can coincide with increased hallucination during complex reasoning.
- Approach: Baichuan-M3 is designed to unify clinical inquiry with reliable decision-making through proactive information acquisition, coherent reasoning trajectories, and adaptive hallucination suppression.Its training framework combines TaskRL, policy distillation, segmented pipeline reinforcement learning, dynamic rubric evolution, and targeted hallucination-suppression objectives.
- Implications: The system aims to move LLMs from passive information retrieval toward active decision-support partnerships that elicit missing data while maintaining diagnostic logic.The contribution explicitly targets hallucination via assumption in open-ended clinical scenarios.
2 Training Infrastructure
Baichuan-M3’s training infrastructure combines patient simulation, rubric-based evaluation, fact-aware verification, and staged optimization to support reliable long-horizon medical training.
- Training infrastructure: The infrastructure integrates a patient simulation environment, verification system, and progressive multi-stage training pipeline.These components provide interaction signals and optimization support for long-horizon medical training.
- Patient simulation: Patient simulation combines passive consultations with interruption-injected interactions to model uncertainty, abrupt questions, and interaction noise.Passive Interaction Mode samples at 75%, while Interruption-Injected Mode samples at 25% and uses both end-of-turn and mid-turn questions.
- Patient simulation: The simulator uses asymmetric visibility for injected snippets because exposing them to the patient simulator can cause instability and deviation from passive-response behavior.The snippet is visible only to the physician agent, which answers the question and continues the consultation while the simulator responds passively.
- Verification: The Verify System separates structural clinical compliance from factual medical validity through parallel Rubric Verifier and Fact Verifier streams.The Rubric Verifier assesses guideline adherence, while the Fact Verifier checks atomic claims against biological and medical knowledge.
- Reward design: Rubric rewards aggregate binary clause decisions with signed weights and min-max normalization, keeping reward scales comparable across rubric configurations.Positive weights denote rewards, negative weights denote penalties, and integer weights encode relative clause importance.
- Verification: Fact-aware verification decomposes responses into independently verifiable atomic claims and validates them against authoritative sources using a search-augmented agent.The verification module runs in parallel with the Rubric Verifier to minimize latency.
3 Task-specific Training Methods
The paper uses task-specific training methods for two medical capabilities: Deep Clinical Consultation and Credible Healthcare Advisory. The consultation pipeline gathers and reasons over staged clinical information, while SPAR and fact-aware reinforcement learning target local validity, factuality, and informative advice.
- Medical practice requires active symptom elicitation, deductive diagnosis, and factual adherence in advisory services, which a monolithic training approach may not balance.
- Deep Clinical Consultation: Deep Clinical Consultation models consultation as a structured, auditable information-production process that collects pivotal clinical data during brief interactions.
- Deep Clinical Consultation: The Segmented Pipeline RL framework decomposes complex clinical tasks into stages and uses accumulated context to connect successive responses and instructions.
- Deep Clinical Consultation: Quality-gated transitions extend only trajectories whose stage-specific quality scores meet an acceptance threshold, pruning error paths from the training curriculum.
- Deep Clinical Consultation: SPAR applies fine-grained step penalties and decoupled advantage estimation to address reward hacking, fragmented logic, and poor local credit assignment in long-horizon interviews.
- Credible Healthcare Advisory: Credible Healthcare Advisory combines dynamic rubric evolution with Fact-Aware Reinforcement Learning to suppress unfaithful hallucinations while preserving detailed medical counsel.
4 Evaluation
The evaluation benchmarks Baichuan-M3 on dynamic clinical workflow simulation and broad medical reasoning against generalist, medical-specific, prior-generation, and physician baselines. It reports leading performance across ScanBench, HealthBench, and hallucination-focused evaluations.
- ScanBench results: Baichuan-M3 leads all three ScanBench stations: Clinical Inquiry 74.9, Laboratory Testing 72.1, and final diagnosis 74.4.Its Clinical Inquiry score exceeds GPT-5.2-High by 12.4 points and the human baseline by over 20 points.
- Inquiry analysis: Baichuan-M3 is the sole model leading across all four SCAN inquiry dimensions, consistently outperforming state-of-the-art LLMs and human experts.The reported dimensions include Safety Stratification, Association & Inquiry, Clarity Matters, and Normative Protocol.
- Dynamic consultation: Across dialogue turns, models converge on symptom clarification around 0.7–0.8, while Baichuan-M3 continues improving in Association & Inquiry and gains nearly a twofold advantage at longer dialogues.Its Safety Stratification score rises to about 0.7 as evidence accumulates.
- HealthBench results: On HealthBench Total, Baichuan-M3 scores 65.1 versus GPT-5.2-High at 63.3, and on HealthBench Hard it scores 44.4 versus 42.0.It also records a 3.5% Hallucination Rate, the lowest reported rate in this comparison.
- Hallucination control: Fact-Aware RL preserves HealthBench Score performance at 65.1 versus 66.2 without reinforcement learning while reducing refuted-response and uncertainty rates by approximately 50%.The reported results frame adaptive weighting as mitigating the trade-off between safety constraints and model utility.
5 Inference Optimization
The paper introduces inference optimizations for Baichuan-M3 that target throughput and memory use. Gated Eagle-3 improves speculative decoding, while self-generated calibration supports near-lossless INT4 quantization.
- Speculative decoding: Standard Eagle-3 speculative decoding can suffer lower candidate acceptance because target-model hidden states overwhelm the lightweight draft model.The paper identifies this representation mismatch as a source of limited achievable speedup.
- Speculative decoding: Gated Eagle-3 adds a Gated-Attention module to regulate information injected into the lightweight Eagle-3 draft model.The gate provides dimension-wise control, emphasizing salient features while suppressing redundant or noisy components.
- Speculative decoding: 12%: Gated Eagle-3 improves average throughput over the Eagle-3 base, alongside an average acceptance-length improvement of 0.31.Speculative decoding verifies proposed candidate tokens in parallel and commits the longest verified prefix.
- Quantization: Self-generated calibration promotes uniform expert coverage to address biased INT4 calibration in sparsely activated Mixture-of-Experts models.The approach uses multi-domain prompts and BF16-generated responses to reduce distributional discrepancies.
- Quantization: INT4-quantized Baichuan-M3 achieves near-lossless performance relative to BF16 on mainstream benchmarks, supporting deployment of large-scale sparse models.The paper presents this strategy as reducing memory requirements without compromising accuracy.
6 Conclusion
Baichuan-M3 models clinical inquiry as part of reliable medical decision-making through workflow-aligned training. The resulting system performs strongly on factual and process-oriented benchmarks.
- Contribution: Baichuan-M3 unifies clinical inquiry with reliable medical decision-making by explicitly modeling clinical workflow.Its stated capabilities are proactive information acquisition, coherent multi-step reasoning, and hallucination suppression.
- Training: The training paradigm combines task-specific reinforcement learning with multi-teacher distillation.The paper describes this as a three-stage training paradigm.
- Conclusion: Baichuan-M3 achieves strong performance on HealthBench, HealthBench-Hallu, and the OSCE-style ScanBench.The authors present workflow-aligned optimization as an effective approach for advancing clinical-grade medical LLMs.
7 Limitation and Future Work
Baichuan-M3 is currently scoped to episodic, text-based clinical scenarios and does not fully cover longitudinal, multimodal, or ultra-long-horizon care. The paper identifies evidence grounding and rare high-risk errors as remaining challenges.
- Limitations: Baichuan-M3 does not fully capture longitudinal disease management, multimodal clinical signals, or ultra-long-horizon reasoning across patient trajectories.The stated scope boundary is episodic, text-based clinical scenarios.
- Future work: Rare high-risk errors and limited explicit grounding in evidence-based sources remain open challenges despite substantially improved hallucination control.Future work targets multimodal inputs, long-context optimization, evidence retrieval, safety constraints, and environment-based reinforcement learning.
8 Contribution
The contribution section lists the paper’s contributors alphabetically by first name and marks former team members with an asterisk.
- Contributors are listed alphabetically according to their first names.
- An asterisk marks contributors who are no longer part of the team.
- The listed contributors include Chengfeng Dou, Fan Yang, Fei Li, Jiyuan Jia, Qiang Ju, Shuai Wang, Tianpeng Li, and Tianpeng Li.
A Appendix
The appendix reports supplementary ablations of SPAR and related training components, including evaluation protocols, reward strategies, and multi-turn consultation results.
- The appendix provides supplementary experiments and detailed analyses to validate the proposed methods.
- The appendix also includes analyses of Clip-Forward-KL, Fact-Aware RL, Gated Eagle-3, and associated training and evaluation prompts.
- SPAR ablation: The SPAR ablation compares a Baichuan-M3 backbone, global-reward GRPO, global repetition penalties, and step-level rewards.
- SPAR ablation: The evaluation uses normalized Repeat Score, Logical Score, Rubrics Score, and Average Turns metrics, with the first two derived from GPT-5 dialogue assessments.
- SPAR ablation: Global rewards improve Rubrics Score but increase redundant inquiries, while global repetition penalties reduce redundancy at the cost of sharply lower Logical Score.
- SPAR ablation: SPAR reduces repetition while preserving logical coherence, allowing more critical medical information to be extracted within fewer consultation turns.
A.2.1 Evaluation of Distilled Claim Extraction Models
This section evaluates smaller SFT claim extractors as efficient substitutes for GPT-5 in online reinforcement learning, emphasizing extraction fidelity and deployment cost.
- Using GPT-5 as the reference extractor is computationally prohibitive for online reinforcement learning.
- Smaller models are fine-tuned with SFT to serve as efficient claim extractors during online RL.
- Extraction fidelity is evaluated by recall, SFT Exclusivity Rate, and GPT Exclusivity Rate relative to GPT-5.
- 72.80% recall is achieved by the 8B model versus 30.45% for the untuned baseline.
- The 32B variant provides marginal gains, but its higher deployment cost is not justified for the online RL pipeline.
- The SFT-8B model is adopted as the claim extractor to balance extraction fidelity with practical deployment constraints.
A.2.2 Ablation Study on Reward Components
The reward-component ablations examine trade-offs between reasoning capability, factual reliability, expert fusion, and speculative decoding efficiency across medical and general benchmarks.
- Reward components: The reward-shaping experiments compare task-only optimization without Fact-Aware RL against a baseline using static hallucination penalties.
- Reward components: The unconstrained model reaches approximately 0.68 on HealthBench but its hallucination rate increases to 0.08.
- Reward components: Static penalties correct hallucination drift but overcompensate, while aggressive suppression severely degrades reasoning capability through penalty-induced conservatism.
- Reward components: Denoise & Reweight reduces hallucinations to approximately 0.035 while maintaining reasoning scores near 0.665.
- Reward components: Its dual-protection design combines marginal noise filtering and competence-based gating to preserve factual correctness and medical utility.
- Offline expert fusion: Clip-Forward-KL preserves inquiry performance while improving healthcare benchmarks by imposing a one-sided lower-bound constraint on teacher-supported actions.
- Offline expert fusion: The offline fusion comparison uses identical initialization, datasets, and training configurations across inquiry and healthcare data.
- Speculative decoding: Gated Eagle-3 improves average acceptance length by 0.31 over Eagle-3 Base under matched training data and protocols.