Baichuan-M4 - A medical augmentation model jointly developed by Baichuan Intelligence and Tsinghua University - AiBoss

What is Baichuan-M4?

Baichuan-M4 is a next-generation medical augmentation model jointly developed by Baichuan Intelligence and Tsinghua University. It simultaneously ranks first in the world on three authoritative HealthBench benchmarks: Comprehensive, Hard, and Professional, with a hallucination rate as low as 3.3%, the lowest in the industry. Baichuan-M4 breaks through the limitations of passive responses in general large-scale models, focusing on four core clinical capabilities: in-depth consultation, full-process disease memory, evidence anchoring, and autonomous agent scheduling, enabling AI to truly move from simply answering questions to effectively diagnosing illnesses.

Main functions of Baichuan-M4

In-depth proactive consultationThe system simulates multiple rounds of questioning by clinicians, guiding patients to provide more details about their symptoms and prioritizing the identification of critical and severe cases, rather than passively waiting for complete information.
Full course of illness memoryBy integrating historical medical records, multiple rounds of consultations, laboratory trends, and medication feedback, we can continuously gain a complete understanding of the patient's medical history through numerous conversations.
anchoring evidenceEach medical conclusion generated corresponds precisely to a specific paragraph in an authoritative paper or guideline, ensuring traceability and verifiability.
Agent schedulingBaichuan-Harness autonomously decides when to ask follow-up questions, retrieve or review medical history, and processes complex subtasks in parallel.
Safety constraintsReal-time interception of unauthorized tool calls, unauthorized data access, and operations that do not conform to clinical standards.

Technical Principles of Baichuan-M4

OSCE-based dynamic consultationDrawing on objective structured clinical examination methods used in medical education, we collaborated with over 150 frontline physicians to develop the SCAN-bench assessment system. The model simulates the real patient consultation process through multiple rounds of dynamic interaction, proactively inquiring about the nature and causes of symptoms to gradually narrow down the diagnostic scope and avoid skipping crucial medical history in order to reach a quick conclusion.
Long-term contextual and full-course memoryBreaking through the limitations of single-turn dialogue memory, this model employs a long-context clinical memory mechanism to continuously integrate structured medical records, previous consultation summaries, examination results, and medication feedback. Throughout multiple interactions across timelines, the model maintains a comprehensive understanding of the patient's identity, past illnesses, and changes in key indicators, providing a personalized data foundation for precision medicine.
Six-source evidence-based practice and evidence anchoringBased on the "six-source evidence-based" paradigm, only authoritative medical sources are retrieved, excluding those crawled from open networks. Guidelines, expert consensus, and real-world treatment processes are broken down into over 1000 standardized clinical pathway units, covering more than 200 diseases. The model output must be precisely anchored to specific paragraphs of the original literature, not just the reference number, ensuring a citation accuracy of 90.0%.
Baichuan-Harness Agent ArchitectureAs the central nervous system of the medical intelligence agent, it autonomously schedules the invocation of the three major modules: consultation, memory, and evidence-based medicine. When faced with heavy tasks, it breaks them down into sub-tasks for parallel processing, reducing the load on the main agent context; at the same time, it has a built-in real-time security guard to block unauthorized tool calls and data access, and supports the iterative return of difficult online cases.

How to use Baichuan-M4

Initial description of symptomsUsers can describe their current physical discomfort or upload test results through the Baichuan Smart Product portal.
Proactive multi-round follow-up questionsM4 automatically initiates targeted follow-up questions to guide users to supplement key information such as symptom location, duration, triggers, and past medical history.
Generate a medical consultation cardAfter completing the information collection, the model organizes the medical history and symptoms into a structured consultation card and provides preliminary medical advice.
Continuous follow-up managementUsers can add new symptoms or test results at any time, and M4 continuously tracks the evolution of the disease based on the memory of the entire course of the disease.

Baichuan-M4's core advantages

Leading in all aspects of the evaluationIt ranked first in the world across three HealthBench benchmarks with a combined score of 68.6, leading the second-place GPT-5.5 by more than 10 points.
Lowest rate of hallucinations in the industryThe rate of factual hallucinations was only 3.3%, significantly lower than that of GPT-5.5 (3.8%) and DeepSeek-V4-Pro (9.8%).
Leading by a significant margin in consultation capabilitiesThe initial SCAN-bench score was 79.0, and the follow-up score was 74.7, both significantly better than the mainstream general large model.
The evidence-based accuracy is extremely high.The Baichuan-EBM citation accuracy reached 90.0, far exceeding GPT-5.5 (54.7).
Long memory spanClinical memory with long context scored 86.9 points, an improvement of 21.1 points compared to the previous generation M3.
Clinically applicableMore than 1,000 clinical pathway units covering over 200 diseases, all verified by senior experts.

Baichuan-M4 project address

arXiv technical paper: https://arxiv.org/pdf/2606.08982

Comparison of Baichuan-M4 with similar competing products

Comparison Dimensions	Baichuan-M4	GPT-5.5
HealthBench Comprehensive	68.6 (Number 1 in the world)	58.4
HealthBench Hard	49.7	33.8
HealthBench Prof	55.1	51.8
Hallucination rate	3.3% (lowest in the industry)	3.8%
SCAN-bench initial diagnosis	79.0	68.8
SCAN-bench follow-up appointment	74.7	67.7
Long-context clinical memory	86.9	81.7
Evidence-based citation precision	90.0	54.7
Consultation mode	Native in-depth proactive questioning, simulating multiple rounds of questioning by a clinician.	Relying on role-playing prompts can lead to hasty conclusions.
Memory mechanism	Full course of illness memory, integrating medical records and follow-ups across timelines	Limited contextual memory; long disease course leads to forgetting of early information.
Evidence tracing	Precisely anchored to specific paragraphs in the paper/guideline	Reference-level citations, insufficient paragraph-level precision
Architecture Design	Baichuan-Harness Agent Autonomous Orchestration and Scheduling	External manual workflow orchestration and multi-module collaboration are required.
Clinical pathway coverage	1000+ standardized pathway units, 200+ diseases	No native clinical pathway system

Application scenarios of Baichuan-M4

Routine health consultationWhen users experience physical discomfort, they can obtain preliminary assessments and medical advice through multiple rounds of follow-up questions, thus avoiding blind medical treatment.
Long-term management of chronic diseasesIt continuously records medication feedback and indicator changes in patients with chronic diseases such as hypertension and diabetes, and provides personalized follow-up reminders.
Pre-diagnosis consultationPatients can complete a symptom review before registering at the hospital, generating a structured consultation card to improve the efficiency of face-to-face consultations.
Cross-regional family careChildren can remotely monitor their parents' health, and the model, combined with long-term records, can identify hidden risks such as early-stage heart failure.
Medical Education and TrainingBased on the OSCE method, dynamic interaction provides medical students with standardized and reusable clinical thinking training.