PaperlayerView exact revision on arXiv

Source-linked AI summary

Baichuan-M2: Scaling Medical Capability with Large Verifier System

M2 Team, Chengfeng Dou, Chong Liu, Fan Yang, Fei Li, Jiyuan Jia, Mingyang Chen, Qiang Ju, Shuai Wang, Shunya Dang, Tianpeng Li, Xiangrong Zeng, Yijie Zhou, Chenzheng Zhu, Da Pan, Fei Deng, Guangwei Ai, Guosheng Dong, Hongda Zhang, Jinyang Tai, Jixiang Hong, Kai Lu, Linzhuang Sun, Peidong Guo, Qian Ma, Rihui Xin, Shihui Yang, Shusen Zhang, Yichuan Mo, Zheng Liang, Zhishou Zhang, Hengfu Cui, Zuyi Zhu, Xiaochuan Wang

arXiv:2509.02208v1cs.LGcs.AI

TL;DR

Static medical benchmarks do not capture the dynamic, interactive nature of clinical consultations, leaving a gap between benchmark performance and real-world clinical decision-making. Baichuan-M2 uses interactive reinforcement learning with simulated patients and dynamically generated clinical rubrics. On HealthBench, it outperformed all other open-source models and most advanced closed-source counterparts, scoring above 32 on HealthBench Hard, a level previously reached only by GPT-5.

  • Problem

    Static medical benchmarks do not capture dynamic clinical decision-making, leaving a gap between exam performance and utility in real-world consultations.

    Source: §1 · p. 1

  • Method

    The verifier system uses patient simulations and dynamically generated multidimensional clinical rubrics for interactive reinforcement learning in simulated clinical environments.

    Source: §1 · p. 2

  • Results

    Above 32 on HealthBench Hard, Baichuan-M2 outperformed all other open-source models and most advanced closed-source counterparts, a level previously reached only by GPT-5.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results underscore the critical role of robust validation systems in integrating model capabilities with practical applications.

    Source: §1 · p. 2

    Baichuan-M2 may still exhibit response hallucinations and insufficient reasoning stability in certain edge cases.

    Source: §7 · p. 17

Abstract

from arXiv · show

As large language models (LLMs) advance in conversational and reasoning capabilities, their practical application in healthcare has become a critical research focus. However, there is a notable gap between the performance of medical LLMs on static benchmarks such as USMLE and their utility in real-world clinical decision-making. This discrepancy arises because traditional exams fail to capture the dynamic, interactive nature of medical consultations. To address this challenge, we introduce a novel dynamic verification framework that moves beyond static answer verifier, establishing a large-scale, high-fidelity interactive reinforcement learning system. Our framework comprises two key components: a Patient Simulator that creates realistic clinical environments using de-identified medical records, and a Clinical Rubrics Generator that dynamically produces multi-dimensional evaluation metrics. Building on this foundation, we develop Baichuan-M2, a 32B-parameter medical augmented reasoning model trained through a multi-stage reinforcement learning strategy with an improved Group Relative Policy Optimization (GRPO) algorithm. Evaluated on HealthBench, Baichuan-M2 outperforms all other open-source models and most advanced closed-source counterparts, achieving a score above 32 on the challenging HealthBench Hard benchmark-previously exceeded only by GPT-5. Our work demonstrates that robust dynamic verifier system is essential for aligning LLM capabilities with practical clinical applications, establishing a new Pareto front in the performance-parameter trade-off for medical AI deployment.

1 Introduction

The introduction argues that static medical exams miss the interactive demands of clinical practice and presents a dynamic verifier and multi-stage reinforcement learning as foundations for Baichuan-M2.

1 Introduction

  • Static medical benchmarks such as the USMLE can miss clinical decision-making ability because consultations involve incomplete information, iterative diagnostic exploration, and nuanced communication.The paper identifies this mismatch between exam performance and practical clinical decision-making as a limitation of conventional evaluation.Source: §1 · p. 1
  • The dynamic verifier replaces static answer checking with simulated clinical practice, using patient simulation and real-time clinical rubrics to evaluate model performance.The Patient Simulator draws on desensitized medical records and doctor-patient conversations to represent diverse backgrounds and personalities; the Clinical Rubrics Generator scores diagnostic accuracy, consultation logic, treatment plans, empathy, and medical ethics.Source: §1 · p. 2
  • Baichuan-M2 combines medical mid-training, rejection-sampled supervised fine-tuning, and multi-stage reinforcement learning with improved GRPO to develop medical reasoning and interaction while maintaining general capabilities.The staged approach decomposes complex reinforcement learning into a controllable hierarchy.Source: §1 · p. 2
  • 32B-parameter Baichuan-M2 outperformed all other open-source models, including gpt-oss-120B, and most advanced closed-source counterparts on HealthBench, scoring above 32 on HealthBench Hard, a level previously reached only by GPT-5.The paper presents these results as evidence that robust validation helps connect model capabilities with practical applications.Source: §1 · p. 2
  • The authors characterize Baichuan-M2 as an open-source model with top-tier performance at lower deployment cost, setting a new performance-parameter Pareto front for medical AI.They suggest this efficiency could make advanced medical AI deployment more feasible in resource-constrained healthcare settings.Source: §1 · p. 2

2 Verifier System

The section argues that static medical answer verification misses the dynamic, partially observable, multi-turn nature of clinical practice. Baichuan-M2 instead uses an interactive reinforcement-learning environment with virtual patients and dynamically generated clinical rubrics.

Motivation

  • Static medical question-bank verifiers miss dynamic clinical judgment and may limit generalization and practical performance.Clinical practice is described as a partially observable, multi-turn decision process involving clinical experience, communication, and ethical considerations.Source: §2 · p. 2

System design

  • The verifier system combines a patient simulator for realistic, stochastic, continuous interactions with a clinical rubrics generator for dynamic assessment across turns.Together, the modules move training beyond rigid single-turn question answering and enable quantitative evaluation of comprehensive performance.Source: §2 · pp. 2–3, Figure 1 · p. 3

Reinforcement-learning loop

  • In the closed loop, the model interacts with virtual patients and uses dense expert-level feedback to iteratively optimize diagnostic strategies.The paper says this process moves capabilities beyond static knowledge recall toward alignment with senior physicians’ clinical thinking and practical skills.Source: §2 · p. 3

2.1 Patient Simulator

The section presents a high-fidelity interactive patient simulator designed to address limitations of static testing and prior simulators’ weak modeling of psychological, social, and interactional complexity. It combines patient scripts, specialized modules, and turn- and session-level evaluation.

2.1 Patient Simulator

  • Prior simulators inadequately model psychological states, social backgrounds, and dynamic interaction, while high-fidelity simulation must balance scenario diversity with reproducible case behavior.The section describes real encounters as involving information withholding, emotional expression, and culturally mediated communication barriers.Source: §2.1 · p. 3

2.1.1 Patient Scripts

  • Patient scripts combine real-world clinical histories with personality and sociocultural profiles, including MBTI-inspired traits that shape communication and treatment responses.The medical information covers chief complaint, illness history, and past medical history across specialties and population groups.Source: §2.1.1 · pp. 3–4

2.1.2 Modules and Inner-interaction

  • A Termination Gate, Affective Unit, and profile-checking Fact Unit address premature endings, behavioral diversity, information leakage, and factual inconsistency, enabling a smaller model to perform comparably to a large one.The Affective and Fact Units use LLMs; the Affective Unit is trained on synthetic data spanning personalities and sociocultural backgrounds.Source: §2.1.2 · p. 4, Figure 2 · p. 4

2.1.3 Performance of Patient Simulator

  • Simulator evaluation averages turn-level Privacy and Fact Scores and pairs them with a session-level Personification Score equally weighting personality and socio-cultural consistency.Privacy measures avoidance of unnecessary personal disclosures, while Fact Score measures adherence to preset records without fabrication.Source: §2.1.3 · p. 4
  • Our method achieves an optimal diversity-consistency tradeoff with fewer parameters, while personification improvements typically accompany reductions in privacy and fact scores.Figure 3 indicates that the proposed method enhances Personification Score while preserving Privacy Score and Fact Score stability.Source: Figure 3 · p. 5, §2.1.3 · p. 5

2.2 Clinical Rubrics Generator

The section presents a Clinical Rubrics Generator that evaluates clinical reasoning with context-specific, expert-aligned criteria. It describes how prompts and rubrics are built and trained, and reports their consistency with expert annotations.

2.2 Clinical Rubrics Generator

  • Binary answer- or rule-matching rewards miss the dynamic decisions, diagnostic reasoning, treatment planning, and communication involved in comprehensive clinical care.The proposed generative verifier is designed to align AI doctors’ reasoning with expert clinical judgment.Source: §2.2 · p. 5
  • Multidimensional rubrics assess diagnostic accuracy and communication, rely on clinician-validated criteria, and adapt to patient-specific factors and communication styles.These attributes are intended to capture clinical competencies while maintaining consistency with professional standards and best practices.Source: §2.2 · p. 5

2.2.1 Prompt Collection and Processing

  • Prompts draw on medical records, evidence-based knowledge sources, and synthetic clinical scenarios, then undergo deduplication, multidimensional scoring, and filtering for clinical value and challenge.The resulting set is described as wide-ranging, high-quality, and balanced, providing a foundation for rubric production and reinforcement-learning training.Source: §2.2.1 · p. 6

2.2.2 Rubric Construction

  • Experts define core assessment dimensions, select and customize LLM-generated candidate rubrics, assign each an integer weight from -10 to 10, and use curated rubrics as seed data for expansion.The workflow addresses initial rubrics’ tendency to be overly uniform and to omit case-specific core points.Source: §2.2.2 · p. 6

2.2.3 Training of Rubrics Generator

  • A mid-trained base model is trained with supervised fine-tuning and reinforcement learning on medical rubrics, math/code reasoning, and complex instruction-following data.The stated aim is to generate real-time evaluation standards across clinical scenarios while controlling online computational cost.Source: §2.2 · p. 5, §2.2.3 · pp. 6–7

2.2.4 Evaluation of Rubrics Generator

  • 92.7% consistency rate was achieved between model-generated and expert-annotated rubrics.Rubrics counted as consistent when they reflected the same evaluation dimension, not when they matched verbatim; GPT-4.1 served as referee. The evaluation used 100 cases sampled across categories from previously generated prompts.Source: §2.2.4 · p. 7

3 Data and Training

The training pipeline progresses from medical-domain mid-training through supervised fine-tuning and multi-stage reinforcement learning, aiming to build medical knowledge, reasoning, and interactive robustness while retaining general capabilities.

3.1 Mid-Training

  • Mid-training combines medical, general, and mathematical corpora at a 2:2:1 ratio, using structured-reasoning data and KL regularization to retain general capabilities,.The medical corpus draws on textbooks, clinical monographs, drug knowledge bases, guidelines, and de-identified medical records; structured rephrasing is constrained to preserve source knowledge.Source: §3.1 · p. 8

3.2 Supervised Fine-Tuning

  • To prepare for reinforcement learning, supervised fine-tuning adds multi-turn clinical dialogues across consultation, triage, EHR generation, medical RAG, and safety scenarios.The final SFT dataset contained 2 million samples, approximately 20% medical-related, and was trained for 2 epochs at 32K context length.Source: §3.2 · pp. 8–9

3.3 Reinforcement Learning

  • Reinforcement learning uses three phases—rule-based, rubric-based, and multi-turn training—with enhanced GRPO [8] modifications for exploration, response-length variation, and training stability.The modifications eliminate KL divergence, use asymmetric clipping and length-normalized loss, and simplify advantage normalization.Source: §3.3 · p. 9

3.3.1 Rule-based RL

  • Rule-based RL kept AIME performance stable while improving SuperGPQA and MedXQA, with clearer gains in medical reasoning than in knowledge-oriented medical QA.Training filters for definitive answers, agreement with reference answers, reasoning demand, and difficulty appropriate to the previous SFT model.Source: §3.3.1 · p. 10

3.3.2 Rubric-based RL

  • A dynamic length reward encourages shorter, comprehensive medical answers only when group rubric-score P80 exceeds a predefined quality threshold and the individual response ranks within the group’s top 80th percentile.Source: §3.3.2 · p. 11

3.3.3 Multi-turn RL

  • Multi-turn RL repeatedly simulates patient dialogue from stratified de-identified cases, generates context-specific rubrics, and uses them to evaluate and reinforce each next response.Training filters out incoherent dialogue fragments; future work aims to optimize complete sessions for cross-turn planning and goal consistency.Source: §3.3.3 · pp. 11–12
  • Baichuan-M2 achieves state-of-the-art performance across all evaluation choices on HealthBench, including overall, hard, and consensus scores.Figure 6 compares Baichuan-M2 with prevailing open-source models on these three evaluation choices.Source: Figure 6 · p. 12

4 Evaluation

Baichuan-M2 performs strongly on HealthBench, including its challenging Hard subset, and outperforms most evaluated models. It also shows an edge in Chinese clinical evaluations and maintains reliable performance on general benchmarks.

4.1 HealthBench

  • On HealthBench core scenarios, Baichuan-M2 ranked first in Emergency Referrals (74.6), Context Awareness (48.0), Context Seeking (55.8), Communication (68.6), Global Health (57.1), and Completeness (67.2).The HealthBench rubric criteria are partitioned into five axes.Source: §4.1 · p. 12, Figure 9 · p. 13
  • Baichuan-M2 surpassed all evaluated open-source models and most advanced closed-source models on HealthBench and HealthBench Hard, with comparable closed-source performance on Consensus.The closed-source comparison reports a notable improvement on the Hard subset relative to other models.Source: §4.1 · p. 12, Figure 7 · p. 13
  • Baichuan-M2 achieved the best cost-effectiveness ratio among the compared open-source models, combining the highest HealthBench score with relatively small scale.The paper frames this result as advancing the model-parameter versus HealthBench-score Pareto front.Source: §4.1 · p. 12, Figure 8 · p. 13
  • Baichuan-M2 scored 34.7 on HealthBench Hard; only it and GPT-5 (46.2) scored above 32 worldwide.HealthBench Hard comprises 1,000 questions focused on real-world scenarios and practical medical solutions.Source: §4.1 · p. 14
  • In a gestational-diabetes example, Baichuan-M2 gave conservative insulin-adjustment guidance and addressed hypoglycemia, fetal assessment, and diet support, while gpt-oss-120B missed hypoglycemia risk.The passage characterizes Baichuan-M2 as stronger in medical completeness, accuracy, and safety for this case.Source: §4.1 · p. 14

4.2 Comparison in China’s Medical Settings

  • Across five Chinese clinical dimensions, Baichuan-M2 was preferred in 67% of Communication, 45% of Examination, 43% of Diagnosis, 37% of Treatment, and 34% of Safety evaluations.The study assessed 57 complex cases from Chinese hospital MDT sessions without definitive gold-standard answers, using expert evaluation across five dimensions and ten weighted metrics, with safety and accuracy weighted highest.Source: §4.2 · pp. 14–15, Figure 11 · p. 15

4.3 General Capability

  • Beyond medical tasks, Baichuan-M2 was evaluated on AIME24/AIME25, IFEval, CF-Bench, Arena-Hard-V2.0, AlignBench, and WritingBench; the authors report stable, reliable general-scenario performance.The benchmarks cover math and STEM, instruction following, and general capability and alignment.Source: §4.3 · pp. 15–16

5 Inference Optimization

Baichuan-M2’s inference optimizations combine quantization for deployment on consumer-grade hardware with speculative decoding to increase generation speed.

5.1 Post-training Quantization

  • W4A16 and W4A8 quantization achieved nearly lossless accuracy using AutoRound, Hadamard transforms, GPTQ, and QQQ packing.The authors report that calibration-data quality and diversity affect accuracy, and that including some original-model responses improved it.Source: §5.1 · p. 16
  • KV cache storage was reduced with FP8 E4M3 quantization, using a static scaling factor of 1.0 for subsequent experiments.Calibration-derived per-layer scales did not significantly change model accuracy compared with the fixed scale; the static strategy supports compatibility with SGLang and vLLM.Source: §5.1 · p. 16
  • 21,133 tokens was the maximum sequence length under W4A8-KV8 on a single RTX 4090, and the model runs on open-source inference engines without code modifications.This single-request SGLang deployment used a GPU with 24 GB VRAM.Source: §5.1 · p. 16, Table 2 · p. 16

5.2 Speculative Decoding

  • The speculative decoding system used a lightweight Baichuan-M2 draft model and Eagle-3 [52] to propose multiple continuations for parallel verification by the target model.The draft model was trained on medical dialogue, clinical notes, and structured medical knowledge, including synthetic medical responses generated by Baichuan-M2.Source: §5.2 · p. 17
  • 2.17× speedup raised throughput from 41.5 to 89.9 tokens/s with speculative decoding on a single RTX 4090.With a 4096-token prompt, the draft model achieved 73% prediction accuracy and an average accepted length of 3.28 tokens per round.Source: §5.2 · p. 17

6 Conclusion

The paper introduces a dynamic reinforcement-learning validation system using interactive patient simulations and multidimensional clinical assessments, and applies it to build and open-source Baichuan-M2. The 32B-parameter model demonstrates strong clinical reasoning and benchmark performance.

6 Conclusion

  • Above 32 on HealthBench Hard, Baichuan-M2 outperformed all other open-source models and rivaled leading closed-source systems, making it one of only two models worldwide to exceed that score.The model has 32 billion parameters.Source: §6 · p. 17
  • The validation system replaces static benchmarks with interactive patient simulations and multidimensional clinical assessment criteria to model real-world clinical decision-making.The authors describe it as bridging LLM evaluation and clinical practice.Source: §6 · p. 17
  • Baichuan-M2 was trained using domain adaptation and multi-stage reinforcement learning, and the authors open-sourced the model.Source: §6 · p. 17

7 Limitation and Future Work

Baichuan-M2 still faces reliability, capability, and optimization limitations, and the authors outline future work to improve inquiry, multi-turn clinical workflows, and medical knowledge grounding.

Limitations

  • Baichuan-M2 may hallucinate or reason unstably in edge cases, and its performance on HealthBench and other real-world evaluations remains far from saturated.The authors also note that tool calling and external knowledge retrieval are not fully optimized in this version.Source: §7 · p. 17

Future Work

  • The current version emphasizes diagnosis and treatment, while future work will strengthen medical inquiry skills and hallucination mitigation through quantitative assessment and optimization.Source: §7 · p. 17
  • The authors plan multi-turn session reinforcement learning to support comprehensive inquiry and diagnosis across the clinical workflow.Source: §7 · p. 17
  • They also plan to explore medical knowledge grounding through knowledge bases and clinical decision support systems to reduce hallucinations and improve diagnostic accuracy.Source: §7 · p. 17

A Example of Evaluation Prompt for Rubrics

The appendix provides paired evaluation prompts that assess the latest assistant response against positive or negative rubric criteria. Each asks for an explanation and a binary judgment of acceptability or unacceptability.

Example of Evaluation Prompt for Positive Rubrics

  • The positive-rubric prompt evaluates whether the latest assistant response meets an acceptable-response criterion and returns an explanation plus an “acceptable” boolean.The prompt defines placeholders for conversation history and a rubric item expressed in natural language with criteria and point values.Source: A Example of Evaluation Prompt for… · pp. 23–24

Example of Evaluation Prompt for Negative Rubrics

  • The negative-rubric prompt evaluates whether the latest assistant response meets an unacceptable-response criterion and returns an explanation plus an “unacceptable” boolean.Source: A Example of Evaluation Prompt for… · p. 24

B Response Case of HealthBench

The section presents Baichuan-M2’s response to a gestational diabetes case, described as superior in accuracy, relevance, and completeness, alongside a gpt-oss-120b response.

B Response Case of HealthBench

  • Baichuan-M2’s gestational diabetes response is described as superior in accuracy, relevance, and completeness; Figure 13 shows a response by gpt-oss-120b.The supplied captions do not provide further comparative results.Source: Figure 12 · p. 25, Figure 13 · p. 26

Related papers

Loading 2509.02208v1…