Technical Summary: Baichuan-M3

Problem Statement

While Large Language Models (LLMs) are increasingly adopted in healthcare, a critical gap remains between passive question-answering and active, clinical-grade decision support. Existing systems often fail to maintain evidence-grounded reasoning and uncertainty awareness in open-ended clinical interactions. Current approaches tend to bifurcate into two disconnected paradigms:

1. Factuality-focused benchmarks: These measure static consistency but often fail to capture the complexity of multi-turn, long-horizon clinical reasoning.
2. Process-oriented simulations: These assess interactive history-taking (e.g., OSCE-style) but often prioritize conversational fluency over deep diagnostic logic.

This disconnect leads to "inquiry inertia" (lack of agency to elicit missing data) or superficial reasoning. Furthermore, technical bottlenecks hinder progress: heterogeneous training environments cause instability in multi-task fusion; long-horizon interactions exacerbate credit-assignment problems in Reinforcement Learning (RL); and efforts to increase reasoning depth often trigger reward saturation and increased hallucinations as models attempt to satisfy complex logical constraints.

Methodology

Baichuan-M3 addresses these challenges through a specialized three-stage training framework and novel optimization strategies designed to unify clinical inquiry with reliable decision-making.

1. Training Infrastructure

• Patient Simulator: To ensure stability during long-horizon training, the authors employ a "passive-personality" patient simulator with two modes: Passive Interaction (75% probability, simulating initial consultations) and Interruption-Injected (25% probability, simulating mid-consultation interruptions). An asymmetric visibility mechanism ensures the physician agent sees the interruption context while the simulator remains passive.
• Verify System: A dual-stream verification system decouples evaluation to prevent "fluent hallucinations":
  • Rubric Verifier: Assesses structural quality and adherence to clinical guidelines using fine-grained, weighted clauses.
  • Fact Verifier: A search-augmented pipeline that decomposes responses into atomic claims and validates them against authoritative medical sources. To accelerate training, a two-level caching system (exact and semantic matching) reduces external search requests by ~85%.

2. Multi-Stage Training Pipeline

The training process follows a "divide-and-conquer" strategy before unification:

• Stage 1: Task-Specific RL (TaskRL): Independent RL pipelines train specialized "expert" teachers for Clinical Inquiry, Healthcare Consultation, and General Reasoning. This isolates gradient interference and allows models to develop distinct inductive biases.
• Stage 2: Offline Policy Distillation: A student model learns from the offline trajectories of the expert teachers using a Clip-Forward-KL objective. This ensures the student broadly covers high-probability regions of different expert distributions without overfitting to single samples or suffering from mode collapse.
• Stage 3: Multi-Teacher Online Policy Distillation (MOPD): The student model re-enters the online environment, constrained by both ground-truth rewards and multi-teacher priors. Using Reverse KL regularization, the model is driven to select the optimal mode when facing conflicting advice, transitioning from an "imitator" to a "decision-maker."

3. Task-Specific Optimization Methods

• Deep Clinical Consultation:
  • Segmented Pipeline RL: The consultation is modeled as a 4-stage process (Inquiry, Differential Diagnosis, Lab Testing, Diagnosis). A Quality-Gated Transition mechanism prunes error propagation by only extending trajectories that pass stage-specific quality verifiers.
  • SPAR Algorithm: A Step-Penalized Advantage with Relative baseline algorithm introduces fine-grained step-wise penalties. It decouples local penalties from the group baseline, enabling an implicit curriculum that first corrects critical errors (e.g., redundancy) and then refines nuances.
• Credible Healthcare Advisory:
  • Dynamic Rubric Evolution: A "Mine-Verify-Inject" loop identifies high-confidence samples with latent defects and dynamically generates new rubric constraints to target specific failure modes, preventing reward hacking.
  • Fact-Aware RL: Instead of naive penalty mechanisms, the model uses Structured Signal Denoising (weighting claims by semantic density to prevent redundancy dilution) and Dynamic Multi-Objective Aggregation. A soft-gating coefficient modulates hallucination penalties based on task performance, ensuring safety constraints are only fully enforced once the model demonstrates sufficient reasoning competence.

Key Contributions

1. Bridging Inquiry and Reasoning: Baichuan-M3 transforms LLMs from passive retrievers into active decision-support partners capable of proactively eliciting missing data while maintaining rigorous diagnostic logic.
2. Clinical-Process-Aligned Optimization: The training paradigm mirrors professional medical training, utilizing Segmented Pipeline RL and Dynamic Rubric Evolution to align model behavior with distinct clinical stages (inquiry, testing, diagnosis), prioritizing evidence-based reasoning over fluency.
3. Superior Empirical Results: The model achieves state-of-the-art performance on factual reliability and interactive procedural rigor, significantly outperforming leading proprietary models and expert baselines in hallucination suppression and diagnostic accuracy.

Experimental Results

Baichuan-M3 was evaluated against state-of-the-art general LLMs (e.g., GPT-5.2-High, Deepseek-V3.2), medical-specific models, and a human baseline of attending physicians.

• ScanBench (OSCE-style Simulation): Baichuan-M3 achieved top performance across all three dimensions:
  • Clinical Inquiry: 74.9 (outperforming GPT-5.2-High by 12.4 points and the human baseline by >20 points).
  • Laboratory Testing: 72.1.
  • Diagnosis: 74.4.
  • The model demonstrated superior "Safety Stratification" (75.8) and "Association & Inquiry" (72.6), showing robustness in long-horizon dialogues where other models degrade.
• HealthBench:
  • Total Score: 65.1 (surpassing GPT-5.2-High's 63.3).
  • HealthBench-Hard: 44.4 (surpassing GPT-5.2-High's 42.0).
  • Hallucination Rate: 3.5% (the lowest among competitors).
• HealthBench-Hallu: In a fine-grained evaluation of factual integrity, Baichuan-M3 reduced the "Refuted Rate" to 2.45% and "Uncertain Rate" to 2.07%, while maintaining high task performance. Analysis of knowledge probes suggests the model reduces "unfaithful hallucinations" (where the model knows the truth but generates falsehoods) by regulating generation to stay within its internal knowledge boundaries.

Significance and Claims

The paper claims that Baichuan-M3 represents a paradigm shift from static QA to dynamic, workflow-aligned clinical support. By explicitly modeling the physician's systematic workflow and employing a hierarchical training strategy that decouples then unifies specialized competencies, the system overcomes the "inquiry inertia" and "hallucination via assumption" limitations of previous models.

The authors assert that workflow-aligned optimization is a critical pathway for advancing clinical-grade medical LLMs. The integration of Fact-Aware RL and Dynamic Rubric Evolution demonstrates that it is possible to suppress hallucinations without inducing the overly conservative behavior that typically degrades reasoning capabilities. The results suggest that Baichuan-M3 can operate safely in practice, offering a level of diagnostic rigor and safety that approaches or exceeds human expert baselines in simulated environments.

Limitations

The authors acknowledge that Baichuan-M3 is currently limited to episodic, text-based scenarios and does not yet fully capture longitudinal disease management, multimodal signals, or ultra-long-horizon reasoning across patient trajectories. While hallucination control is improved, rare high-risk errors and the need for tighter integration with evidence retrieval remain open challenges for future work.