Baichuan-M4: Clinical LLM for Continuous Care

• Baichuan-M4 is a clinical-grade medical agent system defined by three components: a unified runtime (Baichuan-Harness), a core reasoning model trained with continuous-care RL, and a clinical tool layer.
• The system enforces action constraints, safety checks, and multi-agent coordination through a closed-loop orchestration mechanism, ensuring consistency between training and deployment.
• It demonstrates strong benchmark performance with a +15.9 point improvement over GPT-5.5, a 3.3% hallucination rate, and superior evidence retrieval and long-context memory metrics.

Searching arXiv for the specified paper to ground the article in the published record. Baichuan-M4 is a clinical-grade medical large model developed by Baichuan Intelligence for continuous care rather than single-turn medical question answering. It is presented as a coordinated medical agent system organized around three pillars: Baichuan-Harness as a unified runtime, a core reasoning model trained with a continuous-care reinforcement-learning framework, and a clinical tool layer spanning patient-memory management, authoritative evidence-based retrieval, and multimodal medical perception across documents, X-rays, and dermatology. The system is designed so that reinforcement-learning training and real-world deployment share the same runtime interface, while action constraints, tool use, long-term patient memory, and multi-agent coordination are enforced within the same operational loop (Yang et al., 8 Jun 2026).

1. System definition and architectural scope

Baichuan-M4 is explicitly defined as a medical agent system rather than a standalone conversational model. Its architecture is organized around three components. The first is Baichuan-Harness, which functions as the runtime and orchestration layer. The second is the core reasoning model, described as a continuous-care medical LLM trained under a closed-loop RL framework. The third is the clinical tool layer, which provides patient-memory management, evidence-based retrieval, and multimodal perception.

The paper characterizes this design as addressing longitudinal clinical interaction rather than isolated question answering. That distinction is operationalized through persistent patient state, tool-mediated evidence access, and role-specialized agents. A common misunderstanding would be to treat Baichuan-M4 as a monolithic LLM with auxiliary plugins. The reported architecture instead places the runtime layer in a governing position: during both training and inference, all model outputs pass through Harness, which either executes tool calls or returns a “generate next token” command. This suggests that model behavior is intentionally constrained by a systems layer rather than delegated entirely to unconstrained autoregressive generation.

The same section also defines the principal clinical functions exposed through the tool layer. These include a patient-memory manager with short-term versus long-term memory and progressive disclosure, authoritative retrieval based on a six-layer evidence pyramid with PICO decomposition and RL-guided query reformulation, OCR for complex clinical documents, X-ray report generation with structured Findings and Impression, and a dermatology agent that uses a multi-step evidence chain.

2. Baichuan-Harness and closed-loop orchestration

Baichuan-Harness is the unifying runtime for both RL training and real-world deployment. Its stated role is to keep training and deployment consistent while enforcing action-space constraints, safety guardrails, tool invocation, long-term memory management, and multi-agent coordination. The paper emphasizes closed-loop consistency through the exact same API, summarized as Harness $\rightarrow$ Model $\rightarrow$ Tool $\rightarrow$ Harness, in both RL training and deployment (Yang et al., 8 Jun 2026).

Several orchestration mechanisms are specified. Asynchronous SubAgent dispatch is used to parallelize heavy sub-tasks such as literature retrieval and memory summarization. Dynamic role switching occurs between an “information-gathering SubAgent” and a “diagnosis/main Agent,” thereby isolating specialized workflows. Multidimensional patient memory includes structured EHR, past dialogues, and lab/imaging trend summaries. Real-time safety checks are used to block illegal tool calls or non-compliant behaviors.

This runtime-centered design carries a specific methodological implication. The system is not merely tool-augmented; it is action-constrained at the runtime boundary. A plausible implication is that safety and workflow compliance are treated as first-class properties of the execution environment rather than post hoc filters on free-form outputs. That reading is reinforced by the claim that Harness enforces action constraints and performs safety checks before tool execution or continued generation.

The architecture also formalizes multi-agent interaction. The main agent is responsible for overall care planning and final decision. A consultation SubAgent collects user data and asks follow-up questions in a streaming-isolated channel. Specialized SubAgents handle literature retrieval, evidence-chain summarization, and sensor/image processing. Harness orchestrates asynchronous dispatch and result aggregation across these roles.

3. Continuous-care reinforcement learning framework

The core reasoning model is trained under a continuous-care RL framework formalized as an MDP. The paper defines the state as $s_t \in S$, comprising patient short-term context, patient long-term profile, and tool-chain states. Actions $a_t \in A$ are either generation of the next token in the dialogue response or issuance of a tool call, such as retrieving evidence, querying memory, or processing an image. The transition is written as $s_{t+1}=T(s_t,a_t)$ and described as deterministic unfolding of context plus inserted tool outputs. The reward is written as $r_t = R(s_t,a_t,s_{t+1})$ and is decomposed via SPAR++ into span-level signals.

SPAR++ is the paper’s span-level reward modeling mechanism. Instead of assigning a single score to an entire consultation trajectory, it anchors rewards to clinically critical spans such as asking about chest pain, identifying a red-flag symptom, or citing the correct guideline. The framework also introduces quality gating, under which efficiency bonuses such as turn reduction are accrued only when medical-quality thresholds are met. The conceptual sketch provided is

$$R(\tau) \;=\;\sum_{i=1}^{N_{\rm span}} \bigl[\;r^{\rm qual}_i\;+\;\mathbb{1}(Q_{\rm qual}\ge Q_{\rm th})\;r^{\rm eff}_i\bigr],$$

where $Q_{\rm qual}=\sum_i r^{\rm qual}_i$ is the total quality score.

The framework also incorporates reasoning-path compression, described as a fine-grained penalty on the LLM’s internal chain-of-thought tokens $\ell_t$, pushing the model toward more succinct internal reasoning without changing the final user-visible content. The conceptual form is

$\rightarrow$0

The curriculum schedule is two-stage within RL. Stage 1 covers initial-visit tasks such as symptom recognition and standardized history taking. Stage 2 introduces mixed memory tasks combining structured EHR and unstructured summaries, with emphasis on disease-progression reasoning and dynamic strategy under incomplete information.

For policy optimization, the system uses SAPO + R3. SAPO replaces PPO-style hard clipping with a smoother policy-gradient update intended to reduce oscillation, while R3 route replay and KL-divergence constraints are used to suppress loss peaks and prevent entropy collapse. The canonical policy objective given in the paper is

$\rightarrow$1

4. Memory system, evidence retrieval, and multimodal clinical tools

Long-term patient memory is a central subsystem rather than an optional cache. The paper distinguishes two data structures: a Structured Profile, containing chronic diagnoses, surgeries, key lab/imaging trends, and allergies; and Unstructured Summaries, containing past consultation excerpts. Retrieval follows a progressive disclosure algorithm: by default, only a compact summary of Profile plus relevant short-term context is passed to the LLM, while “deep backtracking” calls can fetch full raw records on demand (Yang et al., 8 Jun 2026).

Memory writing is mediated by a defined update chain: Extract $\rightarrow$2 Assess confidence $\rightarrow$3 User confirm $\rightarrow$4 Write to Profile. The stated purpose is to ensure that only verified facts enter long-term memory. This directly bears on a recurrent concern in longitudinal medical agents, namely that persistent memory can amplify errors if unverified inferences are stored as facts. The described update chain is an explicit countermeasure, although the paper also notes future risks related to memory extraction errors.

The evidence-based retrieval stack is structured around a six-layer evidence pyramid: Primary research, reviews, guidelines, practical cases, public health, and regulatory/RWE. Queries are decomposed with PICO, and query reformulation is RL-guided with multi-dimensional rewards over relevance, authority, timeliness, and PICO fit. This indicates that retrieval quality is not treated as a purely lexical or embedding-matching problem; it is embedded in the same decision-theoretic framework as the rest of the system.

The multimodal clinical tool layer contains three principal components. The OCR tool uses a VLM+LLM two-stage pipeline and is evaluated through unstructured similarity and structured bipartite matching for key-value extraction. The X-ray tool decodes Findings and Impression using a VLM$\rightarrow$5LLM pathway and is benchmarked on IU-Xray. The dermatology tool uses an agentic multi-step decision process over a five-dimensional evidence space consisting of hypothesis generation, verification, visual analogy, morphology, and differential diagnosis, with a tool-calling necessity constraint. Taken together, these components indicate that multimodality is incorporated as task-specialized tooling under runtime control rather than through a single undifferentiated multimodal decoder.

5. Training stages and optimization defaults

Training begins with Stage 0, which is supervised fine-tuning on static clinical QA and multi-turn consultation data. This is followed by two RL curriculum phases. Stage 1 (Curriculum RL Phase 1) focuses on initial-visit workflows, with SPAR++ rewards emphasizing complete history taking and red-flag identification. Stage 2 (Curriculum RL Phase 2) targets memory-intensive scenarios mixing structured and unstructured past data.

The paper reports default hyperparameters for the RL setup. The SPAR++ quality threshold is given as $\rightarrow$6 (example), the compression penalty as $\rightarrow$7, and the KL coefficient in the SAPO objective as $\rightarrow$8 (Yang et al., 8 Jun 2026). The losses are described as cross-entropy for SFT and combined policy gradient $\rightarrow$9 for RL.

These details clarify that the training recipe is staged not only by task type but also by memory burden and consultation structure. A plausible implication is that the curriculum is intended to move from relatively standardized initial-visit behaviors toward longitudinal reasoning under incomplete and heterogeneous patient context. The explicit presence of both reward shaping and policy stabilization mechanisms further suggests that consultation quality, efficiency, and behavioral stability are jointly optimized rather than treated as separable post-training objectives.

6. Evaluation profile, reported results, and stated limitations

The evaluation suite is described as cross-dimensional and includes static medical knowledge and safety, dynamic OSCE-style consultation, long-context clinical memory, evidence-based retrieval, medical document OCR, and multimodal image understanding. On HealthBench Hard, Baichuan-M4 is reported at +15.9 points over GPT-5.5. The hallucination rate is reported as 3.3%, identified as the lowest among all compared models. On Scan-Bench V1 (initial) and Scan-Bench V2 (follow-up), the model achieves the highest score. For long-context memory capacity, the reported result is 86.9 vs. 79.4–81.7 (other models).

For evidence-based retrieval under Baichuan-EBM, the system is placed in the top-right corner of the Core Score versus Citation Precision plot, with Citation Precision: 90.0 (vs. 43.8–55.9). In medical document OCR (297 samples), the reported scores are Unstructured: 0.838 and Structured: 0.914, with 227 high-confidence samples above 0.9 and 269 above 0.8. On chest X-ray reports (IU-Xray), the reported values are CIDEr: 0.1892 and GREEN-LLM: 0.8435. In dermatology (f17k subset, 4 893 cases), the system reports TOP-1 EXACT: 30.78%, TOP-1 CATEGORY: 39.53%, TOP-3 CATEGORY: 54.00%, and TOP-6 CATEGORY: 60.68% (Yang et al., 8 Jun 2026).

The paper’s own synthesis attributes the continuous-care advantage to the combination of multi-agent orchestration, long-term memory, and tool use under a unified RL loop. It also attributes the low hallucination rate and strict compliance with clinical workflows to the combined effect of SPAR++ span-level rewards, Harness constraints, and SAPO stabilization.

The limitations are stated explicitly. Coverage gaps remain for rare diseases and atypical presentations. The system may incur errors in memory extraction or tool outputs, including evidence retrieval, OCR, and image-based diagnosis. The outputs are intended for physician supervision only and are not a replacement for professional judgment or regulatory-approved medical devices. These caveats are essential for interpreting the reported metrics: strong benchmark performance is not presented as sufficient for autonomous clinical deployment.