CORE-Acu: Structured Reasoning Traces and Knowledge Graph Safety Verification for Acupuncture Clinical Decision Support
Authors: Liuyi Xu, Yun Guo, Ming Chen, Zihan Dun, Yining Qian, An-Yang Lu, Shuang Li, Lijun Liu
Organizations: College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · School of Information Science and Engineering, Yanshan University, Qinhuangdao, Hebei 066000, China · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · School of Artificial Intelligence, Beihang University, Beijing 100000, China · Key Laboratory of Bioresource Research and Development of Liaoning Province, Northeastern University, Shenyang 110169, China · College of Life and Health Sciences, Northeastern University, Shenyang 110169, China
Large language models (LLMs) show significant potential for clinical decision support (CDS), yet their black-box nature -- characterized by untraceable reasoning and probabilistic hallucinations -- poses severe challenges in acupuncture, a field demanding rigorous interpretability and safety. To address this, we propose CORE-Acu, a neuro-symbolic framework for acupuncture clinical decision support that integrates Structured Chain-of-Thought (S-CoT) with knowledge graph (KG) safety verification. First, we construct the first acupuncture Structured Reasoning Trace dataset and a schema-constrained fine-tuning framework. By enforcing an explicit causal chain from pattern identification to treatment principles, treatment plans, and acupoint selection, we transform implicit Traditional Chinese Medicine (TCM) reasoning into interpretable generation constraints, mitigating the opacity of LLM-based CDS. Furthermore, we construct a TCM safety knowledge graph and establish a ``Generate--Verify--Revise'' closed-loop inference system based on a Symbolic Veto Mechanism, employing deterministic rules to intercept hallucinations and enforce hard safety boundaries. Finally, we introduce the Lexicon-Matched Entity-Reweighted Loss (LMERL), which corrects terminology drift caused by the frequency--importance mismatch in general optimization by adaptively amplifying gradient contributions of high-risk entities during fine-tuning. Experiments on 1,000 held-out cases demonstrate CORE-Acu's superior entity fidelity and reasoning quality. Crucially, CORE-Acu achieved 0/1,000 observed safety violations (95% CI: 0--0.37%), whereas GPT-4o exhibited an 8.5% violation rate under identical rules. These results establish CORE-Acu as a robust neuro-symbolic framework for acupuncture clinical decision support, guaranteeing both reasoning auditability and strict safety compliance.
Figures & tables
Figure 1: Overall architecture of the CORE-Acu framework. The framework operates in two phases: (1) Phase I (Offline Training): Module 1 constructs the S-CoT corpus to model the clinical causal chain, while Module 2 builds the TCM KG. These inputs guide Module 3 to fine-tune the LLM via LMERL, focusing on safety-critical entities. (2) Phase II (Online Inference): Module 4 implements a neuro-symbolic closed-loop mechanism, leveraging KG constraints from Module 2 to verify prescriptions and trigger bounded self-correction for guaranteed safety.
Figure 2: The Triangulated Neuro-Symbolic Validation Framework of CORE-Acu. The protocol begins with a Symbolic Veto Mechanism that utilizes the TCM KG to deterministically intercept hard safety violations, effectively reducing the Safety Violation Rate. Validated outputs then undergo multidimensional adjudication and expert cognitive verification, which scrutinize the causal reasoning chain to distinguish authentic clinical logic from black-box memorization traps.
Category
Node Entity
Semantics Description
Count
Basic Info
AcuCode
Standard alphanumeric code (e.g., LI4).
361
AcuName
Standard Romanized/Pinyin name.
361
Category
The Meridian or Channel attribution.
14
CategoryType
Functional classification.
28
Clinical
Disease
Western or Traditional disease names.
520
Symptom
Discrete clinical symptoms or complaints.
1,240
Table 1: Schema Definition and Statistics of the Proposed TCM-KG. The graph integrates standardized acupoint attributes with expert-verified clinical reasoning logic.
Parameter
Value / Specification
Hardware Infrastructure
CPU
16 vCPU on AMD EPYC 9654
GPU
NVIDIA GeForce RTX 4090 (24GB)
LoRA Fine-tuning Setup
Base Model
Qwen3-8B
Fine-tuning Type
LoRA
Table 2: Implementation Details of Experimental Setup.
Configuration Item
Value / Specification
Judge Model
GPT-5.2 (via OpenAI API)
Call Date
December 2025
Temperature
0.1
Top_p
1.0
Max Tokens
2048
Scoring Scale
0–50 Total Score (5 dimensions, 0–10 each)
Table 3: Detailed Configuration of the LLM-as-a-Judge Evaluation.
Model
BLEU-4
ROUGE-L
Entity-F1
CORE-Acu (Ours)
0.1898
0.3334
0.4612
GLM-4.6
0.0697
0.1970
0.3511
HuatuoGPT-o1-7B
0.0059
0.1209
0.2161
Qwen3-8B (Base)
0.0814
0.1883
0.3062
Qwen-Max (General)
0.0865
0.1988
0.3087
GPT-4o
0.0722
0.1985
0.2906
Table 4: Automated Evaluation Results on the Held-out Test Set. CORE-Acu achieves the best performance among the evaluated baselines across all metrics. Note: Entity-F1 measures the strict overlap of key medical entities (Acupoints, Syndromes) against the ground truth.
Figure 3: Comparative performance across five clinical dimensions.
Figure 4: Distribution of total scores across 1,000 test cases. The comparison includes CORE-Acu, Qwen-Max, Qwen3-8B, GPT-4o, GLM-4.6, and HuatuoGPT-o1-7B.
α
BLEU-4
ROUGE-L
Entity-P
Entity-R
Entity-F1
1.0 (Baseline)
0.3728
0.4452
0.5492
0.5245
0.5350
1.02
0.3715
0.4453
0.5495
0.5284
0.5368
1.5 (Ours)
0.3879
0.4573
0.5667
0.5425
0.5521
2.0
0.3485
0.4225
0.5430
0.4937
0.5139
3.0
0.3455
0.4206
0.5332
0.4806
0.5034
Table 5: Sensitivity analysis of the keyword weight factor α on the internal validation set.
Model
Reasoning
Diagnosis
Pathology Logic
Prescription
Explicability
CORE-Acu (Ours)
9.53
9.33
8.00
8.00
10.00
GPT-4o
6.00
2.00
2.00
10.00
7.00
GLM-4.6
6.33
6.00
6.00
6.33
7.87
HuatuoGPT
4.00
3.00
3.33
7.00
5.67
Qwen-Max
6.00
6.00
2.00
8.00
2.00
Qwen3-8B
5.00
2.00
2.00
8.00
7.00
Table 6: Qualitative clinical expert evaluation (consensus scores, 0–10; n=4 complex cases). Three senior TCM practitioners assessed each model along five criteria. GPT-4o attains the highest Prescription score, whereas CORE-Acu achieves substantially stronger Pathology Logic and Explicability , indicating better alignment between clinical reasoning and the final prescription.
Model
Total Score
Reasoning
Dialectical
Pathology Logic
Prescription
Explicability
GPT-4o
36.73 ± 2.49
7.54 ± 0.56
7.20 ± 0.65
7.09 ± 0.58
7.24 ± 0.63
7.66 ± 0.47
Qwen-Max
36.88 ± 2.10
7.52 ± 0.54
7.16 ± 0.58
6.99 ± 0.59
7.30 ± 0.56
7.91 ± 0.32
Qwen3-8B
29.06 ± 3.56
5.93 ± 0.79
5.75 ± 0.77
5.62 ± 0.72
5.34 ± 1.14
6.43 ± 0.59
GLM-4.6
42.26 ± 2.12
8.77 ± 0.44
8.34 ± 0.60
8.51 ± 0.62
9.04 ± 0.34
7.60 ± 0.58
HuatuoGPT
20.00 ± 3.15
3.99 ± 0.76
3.92 ± 0.67
3.77 ± 0.69
3.64 ± 0.83
4.68 ± 0.73
CORE-Acu
42.96 ± 2.47
8.93 ± 0.59
8.28 ± 0.65
8.74 ± 0.52
8.16 ± 0.72
8.85 ± 0.55
Table 7: Quantitative scale validation on 1,000 clinical cases. Scores are reported as mean ± SD over 1,000 cases using the LLM-as-a-Judge protocol (GPT-5.2). CORE-Acu attains the best overall performance and maintains consistent gains on reasoning-centric dimensions.
Figure 5: Safety Violation Rate across models (lower is better).
Method
Mech.
SVR (%)
Repair (fix/det)
FB (#)
Qwen3-8B (base)
–
30.0
–
–
CORE-Acu (SFT)
–
4.0
–
–
CORE-Acu (full)
veto + rw
0 (0/1,000)
100% (40/40)
0
Table 8: Ablation of KG veto (1,000 test cases).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Role
Content
System
You are a TCM acupuncture expert. Follow the required output schema strictly.
Instruction
Based on the patient’s symptoms, generate a S-CoT response. Your output must be a single valid JSON object and contain no additional text.
Appendix
Table A.1: System and Instruction Template.
Field
Model Output Content
Complaint
Unilateral temporal (lateral) headache, migraine-like, accompanied by nausea and photophobia.
Diagnosis
∙ Shaoyang Headache (Migraine): Unilateral temporal pain linked to the Gallbladder channel. ∙ Hyperactivity of Liver-Yang: Underlying root cause (to be differentiated from Phlegm-Turbidity).
Pathology
Dysfunction of the Shaoyang pivot causing Qi stagnation; Liver Qi rebellion transforming into Fire and disturbing the clear orifices; potentially accompanied by phlegm-turbidity blocking the collaterals.
Principle
∙ Soothe Shaoyang: Unblock the pivot mechanism. ∙ Calm Liver & Extinguish Wind: Address the root cause of rising Yang. ∙ Unblock Collaterals: Relieve local pain.
Acupoints
• SJ5 (Waiguan) + GB41 (Zulinqi): [Main, Bilateral] Mechanism: The Shaoyang pair. SJ5 dredges the pivot mechanism and clears the head/eyes; GB41 opens the Dai Mai . Together they specifically target lateral migraines and reduce sensitivity to external stimuli. • LR3 (Taichong): [Main, Bilateral] Mechanism: Targeting the Root. Courses Liver Qi and pacifies Liver Yang to treat the underlying hyperactivity and reduce the frequency of emotional triggers. • GB20 (Fengchi): [Assist, Bilateral] Mechanism: Dispels wind and unblocks collaterals; pacifies ascending wind-fire to relieve neck tension and head distension. • GB8 (Shuaigu): [Assist, Affected Side] Mechanism: Local Point. Directly enters the temporal collaterals to course wind and stop pain; specifically improves throbbing sensation and photophobia.
Appendix
Table A.3: A concrete S-CoT output example formatted for readability. This example demonstrates how the model explicates the specific mechanism for each acupoint selection, linking back to the diagnosis.
Model
BLEU Scores
ROUGE Scores
Entity Metrics
B-1
B-2
B-3
B-4
R-1
R-2
R-L
Prec.
Rec.
F1
GPT-4o
0.3470
0.1991
0.1205
0.0722
0.2753
0.0961
0.1985
0.3812
0.2368
0.2906
Qwen-Max
0.3788
0.2259
0.1401
0.0865
0.2949
0.1054
0.1988
0.3678
0.2679
0.3087
Qwen3-8B
0.3735
0.2159
0.1343
0.0814
0.2865
0.1025
0.1883
0.3182
0.2970
0.3062
GLM-4.6
0.3655
0.2061
0.1193
0.0697
0.3257
0.1137
0.1970
0.3480
0.3702
0.3511
HuatuoGPT
0.0331
0.0188
0.0107
0.0059
0.2054
0.0443
0.1209
0.4569
0.1431
0.2161
Appendix
Table B.1: Comprehensive breakdown of automated metrics on the held-out test set.
Component
Content / Specification
Trigger Context
Given generated acupoints A and a detected conflicting tuple (x,r,y) where r∈Runsafe .
System Alert
Safety Verifier Alert: Detected unsafe violation under <RULE_TYPE> .
Target S-CoT Response: [complaint :] Unilateral temporal (lateral) headache, migraine-like [diagnosis :] Shaoyang (lateral/temporal) headache / Migraine; Hyperactivity of liver-yang (to be differentiated). [pathology :] Dysfunction of the Shaoyang pivot, stagnation of liver and gallbladder Qi transforming into fire. [principle :] Soothe Shaoyang; Calm the liver and suppress yang; Unblock collaterals to relieve pain. [acupoints :] SJ5, GB41 : main ; GB20, GB8 : assist .
Appendix
Table B.3: Example of a Training Instance in Alpaca Format (SFT Data Record).
Dimension
Assessment Criteria
Reasoning Validity
Assess if the reasoning chain is coherent and free of internal contradictions.
Dialectical Accuracy
Evaluate whether the TCM syndrome (diagnosis) is correct and reflects the precise nature/location of the condition.
Pathological Logic
Verify if the pathological mechanism explanation aligns with TCM theory.
Prescription Rationality
Assess whether selected acupoints align with therapeutic principles and include critical main points.
Explicability
Evaluate transparency, structural clarity, and professional standardization of the output.
Appendix
Table C.1: Definition of the five clinical evaluation dimensions (Scale: 0–10 per dimension).
Component
Content / Specification
System Role
You are a senior TCM acupuncture expert and professor. You need to compare the answers of multiple candidates and score each candidate based on five specific dimensions.
Scoring Criteria
(Detailed rubric as defined in Section C.1 )
Input Data
The prompt receives a formatted block of answers: === Candidate Answer List === {candidates_text_block}
Output Req.
1) Compare all candidates to ensure scoring consistency; 2) Output JSON only (no Markdown); 3) Keys must match the input names exactly.
Appendix
Table C.2: Structure of the Judge Prompt used for evaluation.
Model
Reas.
Dial.
Path.
Pres.
Expl.
Total
Judge’s Rationale (Translated)
CORE-Acu
10
9
10
9
10
48
The reasoning chain is highly coherent, accurately capturing the root-branch relationship. The combination of main points is rigorous and principles-based.
GLM-4.6
9
9
9
9
8
44
Demonstrates a complete chain from etiology to treatment. Syndrome differentiation is precise, though expanded therapies are slightly generalized.
GPT-4o
8
8
8
8
8
40
The differentiation of phlegm-qi binding is accurate with smooth logic. The prescription aligns well, though additions appear somewhat formulaic.
Qwen-Max
8
7
7
7
8
37
Overall structure is clear, but syndrome boundaries are slightly overlapping. The point significance explanation lacks depth for certain key acupoints.
HuatuoGPT
6
5
5
5
6
27
Syndrome differentiation is overly coarse. The prescription is brief with insufficient justification for specific point selection.
Qwen3-8B
4
4
4
3
5
20
Internal logic is loose. Multiple significant errors in acupoint meridian attributes and anatomical locations were detected.
Appendix
Table C.3: Detailed Evaluation Output from Judge LLM (Sample Case 001). Scores represent five key TCM dimensions (0–10).
Field Key
Data Type
Description
complaint
String
Free-text description of patient symptoms.
diagnosis
List[String]
TCM syndrome differentiation (Bian-Zheng).
pathology
String
Mechanism explanation of the illness.
principle
List[String]
Treatment principles guiding the prescription.
acupoints
List[Object]
Structured acupoint prescription. Each object contains: ∙ code: acupoint ID ∙ name: Pinyin name ∙ side: Specific side ∙ role: Therapeutic role
The integration of Large Language Models (LLMs) into clinical decision support is critically obstructed by their opaque and often unreliable reasoning. In the high-stakes domain of healthcare, correct answers alone are insufficient; clinical practice demands full transparency to ensure patient safety and enable professional accountability. A pervasive and dangerous weakness of current LLMs is their tendency to produce "correct answers through flawed reasoning." This issue is far more than a minor academic flaw; such process errors signal a fundamental lack of robust understanding, making the model prone to broader hallucinations and unpredictable failures when faced with real-world clinical complexity. In this paper, we establish a framework for trustworthy clinical argumentation by adapting the Toulmin model to the diagnostic process. We propose a novel training pipeline: Curriculum Goal-Conditioned Learning (CGCL), designed to progressively train LLM to generate diagnostic arguments that explicitly follow this Toulmin structure. CGCL's progressive three-stage curriculum systematically builds a solid clinical argument: (1) extracting facts and generating differential diagnoses; (2) justifying a core hypothesis while rebutting alternatives; and (3) synthesizing the analysis into a final, qualified conclusion. We validate CGCL using T-Eval, a quantitative framework measuring the integrity of the diagnosis reasoning. Experiments show that our method achieves diagnostic accuracy and reasoning quality comparable to resource-intensive Reinforcement Learning (RL) methods, while offering a more stable and efficient training pipeline.
Chen Zhan, Xiaoyu Tan, Gengchen Ma +3
School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China. · Tencent Youtu Lab, Shanghai, 200232, China.
Large language models (LLMs) can generate clinical narratives that are insufficiently grounded in patient-specific evidence. In traditional Chinese medicine (TCM), errors can propagate from etiology and pathogenesis through syndrome diagnosis and treatment principles to prescription generation. We developed TCMClinicalReason-Bench using 2,000 multicenter electronic health record cases to distinguish case-grounded responses from fluent but unsupported diagnostic and therapeutic conclusions. Five general-purpose and two TCM-specific LLMs were evaluated in zero-shot settings. An evidence-constrained rubric assessed seven diagnostic and therapeutic components and three cross-block relations, allowing case-supported alternatives. Qwen3.7-Plus with TCM retrieval served as the automated judge, alongside parallel blinded ratings by five senior TCM clinicians on a 600-case subset. Structural completeness was nearly saturated (99.3-100.0%), but normalized content scores ranged from 40.7% to 54.1%. The five general-purpose models averaged 50.0%, versus 41.3% for the two smaller TCM-specific models. Cross-block logic consistency ranged from 60.8% to 66.8% and correlated moderately with content across cases (Pearson's r = 0.515-0.656). Deficits were greatest in prescription generation, prescription analysis, and symptom-guided modification. In judge stress testing on 100 independent cases, perturbation detection rates across the three relations were 57%, 56%, and 31%, with contradictions detected more reliably than omissions. Separating component quality from cross-block consistency localizes failures missed by endpoint and completeness metrics and identifies where clinician oversight remains necessary.
Jirui Dai, Chenkai Zhang, Yan Jia +4
the Department of Computer Science, Johns Hopkins University, Baltimore, USA · the School of Pharmacy, Nanjing University of Chinese Medicine, Nanjing, China · the School of Information Engineering, Huzhou University, Huzhou, China +3
Large language models have advanced medical reasoning, but clinically grounded reasoning in traditional Chinese medicine (TCM) remains challenging because prescription decisions require integrating domain knowledge, syndrome differentiation, treatment principles, herbal prescription composition, and dosage determination. We develop a TCM language model family comprising models with 4 billion, 8 billion, and 14 billion parameters. The artificial intelligence contribution is a staged clinical alignment framework integrating large scale continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL). We further formulate herbal prescription generation as a structured RL problem with prescription-level feedback, enabling direct optimization of clinical decision outputs beyond knowledge reproduction and generic instruction following. The engineering application is TCM clinical decision support, particularly herbal prescription recommendation. We evaluate the models on a public benchmark containing 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks and on a private retrospective dataset of 4,348 spleen and stomach disease prescription cases. The 14 billion parameter reinforcement aligned model achieves a prescription generation F1 of 37.0 on the public benchmark, compared with 32.3 for DeepSeek-R1 with 671 billion total parameters, and a herb-level F1 of 52.8 on the clinical dataset with improved dosage agreement. Reinforcement post-training consistently improves prescription recommendation across all scales, while physician-source analysis shows that expert heterogeneity affects reward stability. These findings demonstrate a scalable pathway from domain grounding and reasoning supervision to clinically oriented prescription decision support.
Rui Hua, Yu Wei, Ziri Jin +15
School of Computer Science & Technology, Beijing Jiaotong University, Beijing, 100044, China · Tianjin Tasly Digital Chinese Medicine Technology Co., Ltd., Tianjin, 300410, China · Tasly Biopharmaceuticals Co., Ltd., Tianjin, 300410, China +8