Pith. sign in

REVIEW 5 major objections 6 minor 66 references

ZhiFangDanTai: Fine-tuning Graph-based Retrieval-Augmented Generation Model for Traditional Chinese Medicine Formula

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A graph-retrieval plus fine-tuning pipeline is claimed to reduce errors and hallucinations in TCM formula generation.

desk verdict Solid GraphRAG+fine-tuning system for TCM with open-sourced data, but the headline proofs are invalid and the evaluation closes the loop with the same model that generated the references. read the letter →

arxiv 2509.05867 v1 pith:XLREVQXX submitted 2025-09-06 cs.CL cs.AI

ZiXuan Zhang , Bowen Hao , Yingjie Li , Hongzhi Yin This is my paper · ORCID
classification cs.CLcs.AI
keywords TraditionalChineseMedicineformulasGraphRAGretrieval-augmentedgenerationLLMfine-tuningdirectpreferenceoptimizationhallucinationreductionknowledgeexplainable
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish a specific recipe for AI-generated Traditional Chinese Medicine formulas: build a knowledge graph from 80,000 classical formula records, cluster it into seven fine-grained topics, retrieve relevant community summaries for a patient's symptoms, and use those summaries to fine-tune a large language model in two stages (supervised fine-tuning then preference optimization). The authors claim this recipe provably lowers both generalization error and hallucination rates compared with fine-tuning alone, and that the full pipeline beats existing LLM-based TCM models on both a collected test set and real-world clinical consultations. This matters because prior systems either returned formulas without explanations, generated detailed but hallucinated explanations, or could not integrate external knowledge cleanly. A working version of the claim would give clinicians and patients a decision-support tool that outputs structured, reasoned formulas at roughly the inference cost of calling an LLM directly.

What carries the argument

The load-bearing machinery is the GraphRAG retrieval pipeline coupled to two-stage fine-tuning. GraphRAG extracts entities and relations from TCM documents, builds a knowledge graph, detects hierarchical Leiden communities aligned to seven explanation facets (disease, recommended formula, herbs, symptoms and population, pulse and tongue, contraindications, preparation), and answers a symptom query by retrieving top-$k$ answers per community and summarizing them into a global context $c$. SFT trains the LLM to predict the reference answer $y$ from $(x,c)$, and DPO aligns the model toward preferred answers $y_w$ and away from rejected answers $y_l$. The carrying identity is the mutual-information gap $I(y;c|x) \ge \gamma$, which converts retrieval quality into a guaranteed reduction in generalization error, plus the preference-strength term $E[\Delta]$ in the DPO bound.

What would settle it

A reader could run a blinded expert review: have TCM clinicians judge whether outputs are clinically sound, and compare ZhiFangDanTai against a fine-tuned baseline using fully human-authored reference answers on the same cases. If the GraphRAG advantage disappears under expert-judged correctness, the claim that it reduces hallucination would be falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery, stated on its own terms, is that a system which retrieves structured TCM knowledge from a graph and then fine-tunes the generator with SFT plus DPO outperforms both retrieval-only and fine-tuning-only alternatives across a battery of standard and TCM-specific metrics, on both a collected corpus and a zero-shot clinical dataset. It claims theoretical backing for this order: the generalization error upper bound becomes $E(\theta_{\mathrm{Final}}) \le E(\theta_{\mathrm{SFT}}) - \gamma/\beta - E[\Delta]/\beta$, where $\gamma$ lower-bounds the mutual information $I(y;c|x)$ contributed by graph retrieval and $\Delta$ is the DPO preference strength; hallucination is bounded by $P_{\mathrm{hall}}(y|x,c) \le \varepsilon + \delta$, and DPO further suppresses rejected outputs to at most $P_{\mathrm{ref}}(y_l|x,c)e^{-\beta^{-1}\Delta}$. In words, retrieval supplies information the model could not otherwise know, and preference alignment drains probability mass from hallucinated candidates. The paper interprets its experimental improvements as the empirical face of these bounds.

Load-bearing premise

The load-bearing premise is that the detailed reference answers used for training and scoring are clinically correct, even though most were produced by an LLM from fifty human seed examples; if those references contain systematic errors or a consistent stylistic bias, the reported gains may measure imitation of that LLM rather than genuine improvement in formula quality.

Editorial extensions

If this is right

  • GraphRAG plus fine-tuning should dominate retrieval-only and fine-tuning-only models on explainable TCM formula generation, with the largest gap on fine-grained facets such as contraindications and tongue/pulse diagnosis.
  • Adding preference alignment after supervised fine-tuning should further reduce the probability that the model emits a rejected or hallucinated answer, by an amount tied to the preference margin $\Delta$.
  • A model trained this way should transfer zero-shot to clinical records outside its training distribution, supporting cold-start use as a decision-support assistant.
  • Skipping a generative-pretraining stage on the collected corpus is not only cheaper but should avoid distributional drift away from the downstream task.
  • The six TCM-specific evaluation metrics should show consistent gains rather than trade-offs, meaning the system improves factual support while retaining formula-compatibility and logical coherence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same recipe could transfer to other knowledge-intensive structured-output domains, such as herb–drug interaction warnings or clinical guideline summaries, whenever reference documents can be organized into a knowledge graph.
  • Editorial inference: because the reference texts were mostly machine-generated, a direct test of the hallucination claim should separate stylistic imitation from clinical fidelity by re-annotating a held-out set with human TCM experts.
  • Editorial inference: the conflict-aware instruction sets suggest a testable extension in which the model is scored on how often it emits explicit warning messages when given contradictory sources.
  • Editorial inference: the theoretical bounds tie hallucination to retrieval quality $\varepsilon$, so a measurable prediction is that deliberately degrading retrieval (for example, using only one community) should raise hallucination rates approximately linearly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes ZhiFangDanTai, a framework that combines Graph-based Retrieval-Augmented Generation (GraphRAG) with supervised fine-tuning (SFT) and direct preference optimization (DPO) for Traditional Chinese Medicine formula generation. The authors construct a fine-grained TCM instruction dataset from 80,000 collected records, use DeepSeek with 50 human-annotated seeds to generate detailed explanations, build a TCM knowledge graph, and fine-tune Llama3.2-7B. They claim novel theoretical proofs showing that GraphRAG plus fine-tuning reduces generalization error and hallucination rates, and they report large improvements over several baselines on both collected and clinical datasets using BLEU/ROUGE plus six TCM-oriented metrics. The paper also reports ablations over top-k, base models, and the inclusion of GPT, and it releases the model and dataset.

Significance. If the claims were correct, the paper would be a practically useful contribution to domain-specific medical LLM systems: it combines structured retrieval with fine-tuning, evaluates on a real clinical dataset in a zero-shot setting, and releases both the model and the instruction dataset. The authors also make an unusual and welcome effort to design TCM-specific evaluation metrics (CCR, CSCR, CCHR, FS, SCR, LR) and to ablate major design choices. However, the two central pillars of the paper are not currently supported. The theoretical section contains multiple unjustified or internally inconsistent steps, including the key Proposition 1, and the empirical evaluation has a closed loop in which DeepSeek-generated outputs serve simultaneously as training targets, preference sources, and evaluation references. As a result, the headline claims of provable generalization improvement and reduced hallucination are not established by the evidence presented.

major comments (5)
  1. [§III-F.1, Proposition 1] The claimed reduction E(θ_GraphRAG+SFT) ≤ E(θ_SFT) − γ/β does not follow from the proof. The derivation first obtains the per-step bound E(θ_{t+1}) ≤ E(θ_t) − ||∇E(θ_t)||²/(2β), and then, even granting the unproved lower bound ||∇E(θ_t)||² ≥ I(y;c|x)², gives E(θ_{t+1}) ≤ E(θ_t) − γ²/(2β). The text then states that 'after integrating over T optimization steps' this becomes E(θ_GraphRAG+SFT) ≤ E(θ_SFT) − γ/β. This changes the exponent of γ from 2 to 1, drops the factor 1/2, and removes the T-dependence without any derivation. Summing the per-step bound under the stated assumptions would give E(θ_T) ≤ E(θ_0) − Tγ²/(2β), not the claimed expression. Because Proposition 1 is the formal basis for the abstract and contributions, this is a load-bearing error.
  2. [§III-F.1, Proposition 1 proof] The assertion ∥∇E(θ_t)∥² ≥ I(y;c|x)² is introduced without proof. The conditional mutual information I(y;c|x) is a property of the data distribution (x,c,y), whereas ∇E(θ_t) is a property of the loss landscape at a particular parameter vector; no link between the two is established. Earlier in the same proof, I(y;c|x) is used as a distributional quantity in the decomposition E[−log P(y|x,c)] = E[−log P(y|x)] − I(y;c|x), while E(θ_SFT) is later interpreted as an expected negative log-likelihood of the model Pθ(y|x,c). The proof thus conflates the true data distribution with the model distribution and effectively assumes the quantity it needs to prove. This gap cannot be filled from standard β-smoothness or step-size arguments.
  3. [§III-F.2, Proposition 2] The derivation of E(θ_DPO) ≤ E(θ_SFT) − E[∆]/β contains several unjustified steps. The inequality β log(Pθ(yw|x,c)/Pref(yw|x,c)) − β log(Pθ(yl|x,c)/Pref(yl|x,c)) ≥ ∆ is asserted as the DPO optimization objective, but the DPO objective maximizes a log-sigmoid of that expression and no such lower bound follows. The subsequent manipulation uses 'DPO decreases the likelihood of y_l' to write Pref(yl)/Pθ(yl) ≥ 1 and then, in the next paragraph, uses 'DPO increases the likelihood of y_w' to obtain the opposite direction; both cannot be used as independent assumptions in the same bound. The 'uniform preference assumption' E[I(y=y_w)] = E[I(y=y_l)] = 0.5 is also stated without justification, and the final sign of E[∆] is never reconciled with ∆ being a positive log-likelihood ratio. Proposition 2 therefore does not constitute a valid proof of the stated error reduction.
  4. [§III-B.1 and §III-C.2] The evaluation loop is closed. The ground-truth detailed explanations are produced by DeepSeek via in-context learning from only 50 human-annotated seeds, and the same DeepSeek-generated content is used simultaneously as the SFT target, as the source of DPO preference pairs (with answers scored by the reference model), and as the reference for BLEU/ROUGE and for the case-study judgments. Under this protocol, high BLEU/ROUGE scores and low 'hallucination' rates can reflect imitation of DeepSeek's output style rather than clinically correct TCM formula generation. The paper does not report human expert evaluation on a held-out sample of the collected dataset, so the empirical superiority claims in Tables IV–VII are not independently verified. This affects the central empirical claim of the paper, not merely its presentation.
  5. [§III-F.4, Proposition 4] The hallucination bound for DPO is invalid as written. The proof obtains Pθ(yl|x,c) ≤ P_ref(yl|x,c) · e^{β^{-1}∆} · (P_ref(yw|x,c)/Pθ(yw|x,c)), and then invokes the approximation Pθ(yw|x,c) ≈ P_ref(yw|x,c) to conclude Pθ(yl|x,c) ≤ P_ref(yl|x,c) · e^{−β^{-1}∆}. Substituting the approximation into the preceding expression gives Pθ(yl|x,c) ≤ P_ref(yl|x,c) · e^{β^{-1}∆}, not the claimed e^{−β^{-1}∆} bound. The sign error is load-bearing because the stated exponential decay of hallucination probability with ∆ is the entire content of Proposition 4.
minor comments (6)
  1. [§III-B.3 and §III-B.4] The definitions of intra-category and inter-category relationships are swapped between Section III-B.3 and Section III-B.4; the two paragraphs should be reconciled.
  2. [Table IX vs. Table IV] Inference FLOPs are reported inconsistently: Table IV lists GraphRAG and ZhiFangDanTai as ∼1.68·10^5T, while Table IX lists the same entries as ∼1.68·10^6T; the source of the order-of-magnitude discrepancy should be clarified.
  3. [References] Reference [22] attributes Qwen2.5 to 'ByterDance'; the model is released by Alibaba, and the reference should be corrected.
  4. [Throughout] There are numerous typographical errors, including 'heal th', 'T raditional', 'Base on', and 'r' in Section III-B.4; a careful proofread is needed.
  5. [§IV-A] The learning rate is reported as 0.1, which is unusually high for LoRA fine-tuning of a 7B model; this should be double-checked against the actual training configuration.
  6. [§III-F.3 and §III-F.4] The notation F, F_x, and F\F_x is introduced without a clear definition of the fact set, and the relationship between ε and retrieval quality is asserted rather than defined.

Circularity Check

3 steps flagged · score 7.0 of 10

The theoretical reduction is an assumed mutual-information threshold restated as a conclusion, and the SFT/DPO/evaluation loop is closed around LLM-generated references.

  1. self definitional [Section III-F.1, Proposition 1 and its proof (Eq. (3))]
    "When I(y;c|x)≥γ, combining GraphRAG with SFT yields a generalization error E(θ_GraphRAG+SFT) bounded by E(θ_GraphRAG+SFT)≤E(θ_SFT)−γ/β. ... E[−logP(y|x,c)] =E[−logP(y|x)]−E[log P(y|x,c)/P(y|x)]. Since E[log P(y|x,c)/P(y|x)]=I(y;c|x), I(y;c|x)≥γ and E[−logP(y|x)]=E(θ_SFT), we have E(θ_GraphRAG+SFT)=E(θ_SFT)−I(y;c|x)≤E(θ_SFT)−γ."

    The claimed reduction in Eq. (3) is exactly the definitional identity that conditioning on c lowers log-loss by the conditional mutual information I(y;c|x). The proof never shows that GraphRAG produces a context with I≥γ; it assumes that threshold as the proposition's antecedent. The bound is therefore the input assumption restated as the conclusion, with the β-scaling and the later γ²/2β→γ/β step asserted without derivation. The substantive claim 'GraphRAG reduces generalization error' is not proved; it is the assumed information gain in new notation.

  2. fitted input called prediction [Section III-B.1, Section III-C.1, and Section IV.A (SFT data and evaluation references)]
    "we employ DeepSeek [12] to extract detailed information for each prescription using its name and ingredients as input. ... we first annotate 50 samples as reference examples, then utilize the model to verify and refine the remaining data. ... The Output content y is the ground-truth answers, which are obtained through human annotation and in-context learning with a LLM."

    These DeepSeek-produced texts are used both as the SFT training targets in Eq. (1) and as the references for BLEU/ROUGE evaluation on the held-out split. A model trained to maximize log P(y|x,c) on this distribution will be rewarded on the test metrics for imitating DeepSeek's output style, so the reported lexical 'improvement' is substantially a measure of fit to the same generator that created the labels. The evaluation is not independent of the training target distribution; only the 50 human seeds and the expert clinical formulas are external anchors.

1 more flagged steps
  1. fitted input called prediction [Section III-C.2 and Section IV.A (DPO preference construction)]
    "we require the LLM π_ref generating two answers and score them according to quality of the answers. We denote the answer with higher score as y_w, and the one with lower score as y_l. ... we require the supervised fine-tuned LLM π_ref to generate 3,000 pairs of answers and score them as {y_w,y_l}, which are then used to execute DPO."

    The DPO loss in Eq. (2) is minimized by increasing πθ(y_w)/πref(y_w) relative to πθ(y_l)/πref(y_l). Since π_ref both generates and scores the preference pairs, the optimization aligns πθ to π_ref's own self-assessment rather than to an external preference signal. The subsequent hallucination bound in Proposition 4, Pθ(y_l)≤P_ref(y_l)e^{−β^{-1}Δ}, is a direct algebraic consequence of this same DPO objective, so the claimed DPO-driven hallucination reduction is defined by the reference model's own preferences, not by independent clinical validation.

full rationale

ZhiFangDanTai's central theoretical claim is not derived from GraphRAG's mechanism: Proposition 1's bound is the conditional-marginal log-loss identity E[−log P(y|x,c)] = E[−log P(y|x)] − I(y;c|x) with the threshold I≥γ assumed as antecedent. The substantive claim that GraphRAG actually supplies that information is never proved, so the 'reduction' is the assumption in new notation. The empirical loop is also closed: DeepSeek produces the fine-grained explanations used as SFT labels and BLEU/ROUGE references, so high lexical similarity to DeepSeek is partially trained in by construction. DPO preferences are generated and scored by π_ref, the very model being aligned, making the reported preference-alignment gains self-referential. The paper does contain some independent elements: the six TCM-oriented metrics (e.g., CCR, CSCR), a clinical dataset whose ground-truth formulas came from TCM experts, and 50 human-annotated seeds. These prevent a score of 10 and give the empirical comparison partial non-circular content. The proof's additional algebraic problems, such as the unexplained γ²/2β to γ/β switch and the asserted gradient-norm lower bound, are correctness defects rather than circularity, so they are not scored as separate circular steps.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The proofs introduce unmeasured quantities (gamma, beta, epsilon, delta) and unjustified inequalities; the empirical pipeline uses LLM-generated references and self-scored preference pairs, creating a closed loop. No new physical entities are invented.

free parameters (7)
  • CSCR weights w_s, w_mi, w_a, w_me = 0.25 each
    Set 'without loss of generality' in Table III; these weights directly determine the reported CSCR score.
  • LoRA rank tau = 12
    Chosen after observing Appendix D, Fig. 4a; performance peaks at tau=12.
  • DPO temperature beta = 0.2
    Chosen in Appendix D, Fig. 4b; the same letter beta is also reused as a Lipschitz constant in the proofs.
  • Retrieval top-k = 3 for main results; 2 for efficiency
    Main tables use 3; Section IV-C.2 says 2 or 3 perform best and 2 is selected for computational efficiency.
  • Chunk size = 512 tokens
    Selected through 'empirical findings' in Section III-B.2 without a sensitivity analysis.
  • Mutual information threshold gamma = not specified
    Introduced in Proposition 1 to make the bound positive; no value or measurement is given.
  • Proof Lipschitz constant beta = not specified
    Assumed in Proposition 1 for beta-smoothness; no finite value is established and it is conflated with the DPO temperature.
assumptions (7)
  • ad hoc to paper DeepSeek-generated explanations, seeded by 50 human-annotated examples, are treated as ground truth for the remaining 80k records.
    Section III-B.1 uses in-context learning with DeepSeek to build the instruction dataset; the same generated texts are used as evaluation references, so clinical correctness is not independently established.
  • ad hoc to paper The conditional mutual information I(y;c|x) is at least gamma for some positive gamma.
    Proposition 1 assumes this lower bound; without it the claimed error reduction vanishes.
  • ad hoc to paper The SFT loss is beta-smooth and the gradient norm squared is at least the squared mutual information.
    Proposition 1 asserts both properties with no derivation; the gradient bound is the load-bearing step.
  • ad hoc to paper Uniform preference assumption with E[I(y=y_w)] = E[I(y=y_l)] = 0.5.
    Proposition 2 uses this assumption, but under it the preference term averages to zero instead of -E[Delta]/beta.
  • ad hoc to paper Approximate fidelity P_theta(y_w|x,c) approximately equals P_ref(y_w|x,c).
    Proposition 4 uses this approximation to change the sign of the hallucination bound.
  • domain assumption Retrieved context covers the relevant facts with probability at least 1-epsilon.
    Proposition 3 assumes retrieval coverage; epsilon is never measured, so the claimed hallucination bound is not instantiated.
  • domain assumption The official TCM databases and the Haodf.com clinical records are accurate and representative.
    The knowledge graph and the zero-shot clinical evaluation rest on these external sources, whose quality is not audited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZhiFangDanTai: Fine-tuning Graph-based Retrieval-Augmented Generation Model for Traditional Chinese Medicine Formula." pith.science (2026). https://pith.science/paper/XLREVQXX

@misc{pith2026250905867,
  author       = {Pith},
  title        = {Pith review of: ZhiFangDanTai: Fine-tuning Graph-based Retrieval-Augmented Generation Model for Traditional Chinese Medicine Formula},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XLREVQXX}},
  note         = {Machine review of arXiv:2509.05867}
}
read the original abstract

Traditional Chinese Medicine (TCM) formulas play a significant role in treating epidemics and complex diseases. Existing models for TCM utilize traditional algorithms or deep learning techniques to analyze formula relationships, yet lack comprehensive results, such as complete formula compositions and detailed explanations. Although recent efforts have used TCM instruction datasets to fine-tune Large Language Models (LLMs) for explainable formula generation, existing datasets lack sufficient details, such as the roles of the formula's sovereign, minister, assistant, courier; efficacy; contraindications; tongue and pulse diagnosis-limiting the depth of model outputs. To address these challenges, we propose ZhiFangDanTai, a framework combining Graph-based Retrieval-Augmented Generation (GraphRAG) with LLM fine-tuning. ZhiFangDanTai uses GraphRAG to retrieve and synthesize structured TCM knowledge into concise summaries, while also constructing an enhanced instruction dataset to improve LLMs' ability to integrate retrieved information. Furthermore, we provide novel theoretical proofs demonstrating that integrating GraphRAG with fine-tuning techniques can reduce generalization error and hallucination rates in the TCM formula task. Experimental results on both collected and clinical datasets demonstrate that ZhiFangDanTai achieves significant improvements over state-of-the-art models. Our model is open-sourced at https://huggingface.co/tczzx6/ZhiFangDanTai1.0.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 60 canonical work pages

  1. [1]

    Proposition 1:GraphRAG+SFT reduce generalization error LetI(y;c|x)be the conditional mutual information between GraphRAG-generated contentc(x,G)and target answery. WhenI(y;c|x)≥γ, combining GraphRAG with SFT yields a generalization errorE(θ GraphRAG+SFT) bounded by E(θ GraphRAG+SFT)≤E(θ SFT)− γ β,(3) whereγrepresents the minimum mutual information thresh-...

    work page

  2. [2]

    Proof:The optimal solution for DPO ensures: Pθ(y|x,c) = 1 Z(x,c) Pref(y|x,c)eβ−1r(x,y), wherer(x,y)is implicit reward

    Proposition 2:Incorporating DPO can further reduce the generalization errorBuilding upon the combined GraphRAG and SFT approach, the incorporation of DPO yields the following generalization error bound for preference- aligned data: E(θ DPO)≤E(θ SFT)− E[∆] β ,(4) where∆ = log Pref(yw|x,c) Pref(yl|x,c) represents preference strength. Proof:The optimal solut...

    work page

  3. [3]

    Proposition 3:Reduced Hallucination in GraphRAG+SFTWhen the retrieved contextc=π θ(x,G∗ x) from GraphRAG covers the relevant facts ofxwith probability P(Fx ⊆c(x,G))≥1−ε, the hallucination probability Phall(y|x,c(x,G))is bounded byε+δ, i.e., Phall(y|x,c(x,G))≤ε+δ,(7) whereεrepresents the training error, reflecting retrieval quality (lowerεindicates better ...

    work page

  4. [4]

    This probability is upper-bounded byP ref(yl|x,c)·e −β−1∆, where ∆ = log Pref(yw|x,c) Pref(yl|x,c)

    Proposition 4:Reduced Hallucination in GraphRAG with SFT and DPOWhen combining GraphRAG with SFT and further optimizing through DPO, the model’s prob- ability of generating hallucinated outputsy l is reduced. This probability is upper-bounded byP ref(yl|x,c)·e −β−1∆, where ∆ = log Pref(yw|x,c) Pref(yl|x,c) . Proof:From the proof of Proposition 2, we have ...

    work page 2022

  5. [5]

    LLaMA [7], Kimi [11], DeepSeek [12], Qwen [22]: Plug-and-play approaches using LLaMA3.2-7B, Kimi-7B, DeepSeek-7B, and Qwen2.5-7B to directly output TCM for- mulas and explanations

    work page

  6. [6]

    The constructed dataset only contains the composition of the formula’s sovereign, minister, assistant, courier based on symptoms

    TCMLLM [17]: A FT approach, which constructs a TCM formula instruction dataset to fine-tune LLaMA and output TCM formulas and explanations. The constructed dataset only contains the composition of the formula’s sovereign, minister, assistant, courier based on symptoms

    work page

  7. [7]

    For efficient similarity search, we utilize FAISS (Facebook AI Similarity Search 10), an embedding matching library implemented in LlamaIndex, to retrieve the top-k results

    RAG [34]: A RAG approach using Llama3.2-7B as the base model. For efficient similarity search, we utilize FAISS (Facebook AI Similarity Search 10), an embedding matching library implemented in LlamaIndex, to retrieve the top-k results. Notably, all following methods involving RAG use the above experimental settings

    work page

  8. [8]

    GraphRAG [10]: A GraphRAG approach using Llama3.2- 7B as the base model. Perform local retrieval in documents based on symptoms and aggregate local retrieval information to obtain global retrieval content, then output TCM formulas and explanations based on symptoms and the global retrieval 10https://github.com/facebookresearch/faiss content. In practice, ...

    work page

Show all 66 references
  1. [9]

    The process includes constructing a TCM formula instruction dataset for fine-tuning, involving both SFT and DPO processes

    RAG+Fine-tuning: A RAG+FT approach using Llama3.2- 7B as the base model and fine-tune it, where the fine- tuning dataset is constructed by RAG. The process includes constructing a TCM formula instruction dataset for fine-tuning, involving both SFT and DPO processes

  2. [10]

    RAG+SFT [36]: A RAG+FT approach using Llama3.2-7B as the base model that only performs SFT during the fine- tuning step, where the fine-tuning dataset is constructed by RAG

  3. [11]

    The framework integrates GraphRAG with SFT and DPO techniques

    ZhiFangDanTai: Our proposed model implements a GraphRAG+FT approach, utilizing Llama3.2-7B as the base model. The framework integrates GraphRAG with SFT and DPO techniques

  4. [12]

    For all retrieval-augmented methods (RAG, GraphRAG, RAG+Fine-tuning, RAG+SFT, GraphRAG+SFT, and ZhiFang- DanTai ), we consistently set top-kas 3

    GraphRAG+SFT: A GraphRAG+FT approach A method that only performs SFT during the fine-tuning step of the Zhi- FangDanTai model. For all retrieval-augmented methods (RAG, GraphRAG, RAG+Fine-tuning, RAG+SFT, GraphRAG+SFT, and ZhiFang- DanTai ), we consistently set top-kas 3. C. E...

  5. [13]

    The results are shown in Table IV

    Overall Results:To explore whether ZhiFangDanTai out- performs the baseline methods in terms of overall perfor- mance (RQ1), we set top-kas 3 and evaluated each model on the test set. The results are shown in Table IV. The results indicate that: 1) ZhiFangDanTai achieves the h...

  6. [14]

    Detailed analysis :To explore what is the impact of choosing different top-kduring the local and global informa- tion retrieval processes (RQ3), we evaluate the models with different top-kvalues, i.e., top-k= 1, 2, and 3, and report the results in Fig. 3. The results show that...

  7. [15]

    From the perspective of risk of overfitting, Goar et al. [40] have pointed out the model’s generalization error is bounded by the Rademacher complexityR N(F), i.e., ϵ(θGeneralization Error)≤R N(F) +O q log(1/δ) N , whereF 12 JOURNAL OF BIOMEDICAL AND HEAL TH INFORMA TICS T ABL...

  8. [16]

    Section III-E) while evaluating competitors through their official platforms

    Model inference analysis :To examine the total inference time encompassing GraphRAG retrieval and answer generation (RQ6), we compare ZhiFangDanTai with DeepSeek and Kimi using all clinical and collected test samples, deploying them via WebUI (C.f. Section III-E) while evaluat...

  9. [17]

    The results, presented in Table VII, indicate that ZhiFangDanTai achieves the highest performance and the lowest hallucination rate

    Clinical Data Performance:To assess ZhiFangDanTai ’s performance on cold-start clinical data relative to baseline methods (RQ7), we conducted a comparative evaluation by directly prompting both ZhiFangDanTai and baseline models to generate TCM formulas from the given clinical ...

  10. [18]

    Research on mining method and application of TCM com- patibility rule based on statistical learning theory.Nanjing University of Chinese Medicine, 2024

    Tang J. Research on mining method and application of TCM com- patibility rule based on statistical learning theory.Nanjing University of Chinese Medicine, 2024

    2024

  11. [19]

    Research on compatibility of traditional Chinese medicine based on big data thinking, 2017

    JiXiang L. Research on compatibility of traditional Chinese medicine based on big data thinking, 2017

    2017

  12. [20]

    Application of complex network method in data mining of traditional Chinese medicine compound, 2021

    Ni M. Application of complex network method in data mining of traditional Chinese medicine compound, 2021

    2021

  13. [21]

    2021;450–455

    Li S, Wang W, He J, KGAPG: Knowledge-aware neural group repre- sentation learning for attentive prescription generation of traditional Chinese medicine, in BIBM’21, pp. 2021;450–455

    2021

  14. [22]

    FordNet: recommending traditional Chinese medicine formula via deep neural network integrating phenotype and molecule

    Zhou W, Yang K, Zeng J, et al. FordNet: recommending traditional Chinese medicine formula via deep neural network integrating phenotype and molecule. Pharmacol Res. 2021;173: 105752

    2021

  15. [23]

    A novel transfer learning model for traditional herbal medicine prescription generation from unstructured resources and knowledge,

    Z. Liu, C. Luo, D. Fu, J., et al. “A novel transfer learning model for traditional herbal medicine prescription generation from unstructured resources and knowledge,” Artif. Intell. Med., pp. 102232, 2022

    2022

  16. [24]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  17. [25]

    ZhipuAI, ChatGLM, 2024

    2024

  18. [26]

    TCM-FTP: Fine-Tuning Large Language Models for Herbal Prescription Prediction

    Zhou, X., Dong, X., Li, C., et al. TCM-FTP: Fine-Tuning Large Language Models for Herbal Prescription Prediction. arXiv preprint arXiv:2407.10510

    arXiv

  19. [27]

    From local to global: A graph rag approach to query-focused summarization

    Edge, D., Trinh, H., Cheng, N., et al. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130

    arXiv

  20. [28]

    https://www.kimi.com/

    Kimi, 2023, MoonShot. https://www.kimi.com/

    2023

  21. [29]

    https://www.deepseek.com/

    DeepSeek-v3, 2024. https://www.deepseek.com/

    2024

  22. [30]

    Liu, et al.The Instruction Tuning of Large Language Models with Multi-Modal Recommendation Instruction

    B, Hao., Yi. Liu, et al.The Instruction Tuning of Large Language Models with Multi-Modal Recommendation Instruction. J. China Univ. Posts Telecommun., 2024, 47(4): 36-43

    2024

  23. [31]

    Lima: Less is more for alignment

    Zhou, C., Liu, P., Xu, P., et al. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36, 2024

    2024

  24. [32]

    Direct preference optimization: Your language model is secretly a reward model

    Rafailov, R., Sharma, A., Mitchell, E. Direct preference optimization: Your language model is secretly a reward model. In NeurlPS’24

  25. [33]

    TCMPR: TCM prescription recommendation based on subnetwork term mapping and deep learning

    Dong, X., Zheng, Y ., Shu, Z., et al. TCMPR: TCM prescription recommendation based on subnetwork term mapping and deep learning. BioMed Research International, 2022(1), 4845726

    2022

  26. [34]

    TCMLLM, https://github.com/2020MEAI/TCMLLM

    Xue, Z., Yang, K., Hao, T., et al. TCMLLM, https://github.com/2020MEAI/TCMLLM

  27. [35]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Lewis, P., Perez, E., Piktus, A., et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurlPS’20, pp.33, 9459-9474

  28. [36]

    Arigraph: Learning knowledge graph world models with episodic memory for llm agents

    Anokhin, P., Semenov, N., Sorokin, A., et al. Arigraph: Learning knowledge graph world models with episodic memory for llm agents. arXiv preprint arXiv:2407.04363, 2024

    2024 arXiv

  29. [37]

    Q., Gu, J

    Yan, S. Q., Gu, J. C., Zhu, Y . et al. Corrective retrieval augmented generation. arXiv preprint arXiv:2401.15884, 2024

    2024 arXiv

  30. [38]

    Best-first beam search

    Meister, C., Vieira, T., Cotterell, R. Best-first beam search. In TACL’20, pp, 8, 795-809

  31. [39]

    https://chat.qwen.ai/

    Qwen2.5, ByterDance, 2024. https://chat.qwen.ai/

    2024

  32. [40]

    ”Research on multimodal social media information popularity prediction based on large language model.” Journal on Communication/Tongxin Xuebao 45.11 (2024)

    Jie, W., et al. ”Research on multimodal social media information popularity prediction based on large language model.” Journal on Communication/Tongxin Xuebao 45.11 (2024)

    2024

  33. [41]

    Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Lowe, R. Training language models to follow instructions with human feedback. In NeurlPS’22, 35, 27730-27744. 14 JOURNAL OF BIOMEDICAL AND HEAL TH INFORMA TICS

  34. [42]

    A., Waltman, L., and Van Eck, N

    Traag, V . A., Waltman, L., and Van Eck, N. J. From Louvain to Lei- den: guaranteeing well-connected communities. Scientific Reports, 2019, 9(1)

    2019

  35. [43]

    Dean, J., & Ghemawat, S. (2008). MapReduce: simplified data processing on large clusters. Communications of the ACM, 51(1), 107-113

    2008

  36. [44]

    Chen, W., Li, Z., Fang, H., Yao, Q., Zhong, C., Hao, J., ... Wei, Z. (2023). A benchmark for automatic medical consultation system: frameworks, tasks and datasets. Bioinformatics, 39(1), btac817

    2023

  37. [45]

    Wang, T., Yang, J., Xiao, Y ., Wang, J., Wang, Y ., Zeng, X., ... Peng, J. (2023). DFinder: a novel end-to-end graph embedding-based method to identify drug–food interactions. Bioinformatics, 39(1), btac837

    2023

  38. [46]

    Lu, D., Zheng, Y ., Yi, X., Hao, J., Zeng, X., Han, L., ... Peng, J. (2025). Identifying potential risk genes for clear cell renal cell car- cinoma with deep reinforcement learning. Nature Communications, 16(1), 3591

    2025

  39. [47]

    KGE-UNIT: toward the unification of molecular interactions prediction based on knowledge graph and multi-task learning on drug discovery

    Zhang, C., Zang, T., Zhao, T. KGE-UNIT: toward the unification of molecular interactions prediction based on knowledge graph and multi-task learning on drug discovery. Briefings in Bioinformatics, 2024, 25(2), bbae043

    2024

  40. [48]

    scMultiGAN: cell-specific imputation for single-cell transcriptomes with multiple deep generative adversarial networks

    Wang, T., Zhao, H., Xu, Y ., Wang, Y ., Shang, X., Peng, J., Xiao, B. scMultiGAN: cell-specific imputation for single-cell transcriptomes with multiple deep generative adversarial networks. Briefings in Bioinformatics, 2023, 24(6), bbad384

    2023

  41. [49]

    J., Ting, D

    Thirunavukarasu, A. J., Ting, D. S. J., Elangovan, K., Gutierrez, L., Tan, T. F., & Ting, D. S. W. Large language models in medicine. Nature medicine, 2023, 29(8), 1930-1940

    2023

  42. [50]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation

    Bao, K., Zhang, J., Zhang, Y ., Wang, W., Feng, F., He, X. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. In Recsys’23, pp. 1007-1014

  43. [51]

    Benchmarking retrieval- augmented generation for medicine

    Xiong, G., Jin, Q., Lu, Z., & Zhang, A. Benchmarking retrieval- augmented generation for medicine. In Findings of ACL’24, pp. 6233-6251

  44. [52]

    Medical Graph RAG: Towards Safe Medical Large Language Model via Graph Retrieval-Augmented Generation

    Wu, J., Zhu, J., & Qi, Y . Medical Graph RAG: Towards Safe Medical Large Language Model via Graph Retrieval-Augmented Generation. arXiv 2024. arXiv preprint arXiv:2408.04187

    2024 arXiv

  45. [53]

    & Aghemo

    Giuffre, M., Pugliese, N., Kresevic, S., Ajcevic, M., Negro, F., Puoti, M., ... & Aghemo. A. Expert-Validated Framework for Large Language Models in HCV Care: A Comparative Analysis of RAG and Fine-Tuning Using Clinical Guidelines. Digestive and Liver Disease, 2025, 57, S13-S14

    2025

  46. [54]

    T., Koh, P.,

    Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W. T., Koh, P., ... & Hajishirzi, H. FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. In EMNLP’23, pp. 12076- 12100

  47. [55]

    & Han, W

    Wei, X., Cui, X., Cheng, N., Wang, X., Zhang, X., Huang, S., ... & Han, W. Zero-shot information extraction via chatting with chatgpt. arXiv, 2023

    2023

  48. [56]

    Lopez, I., Swaminathan, A., Vedula, K., Narayanan, S., Nateghi Haredasht, F., Ma, S. P., ... & Chen, J. H. (2025). Clinical entity augmented retrieval for clinical information extraction. npj Digital Medicine, 8(1), 45

    2025

  49. [57]

    Goar, V ., & Yadav, N. S. (2024). Foundations of machine learning. In Intelligent Optimization Techniques for Business Analytics, 2024, pp. 25-48

    2024

  50. [58]

    Hao, B., Zhang, J., Yin, H., Li, C., & Chen, H. (2021). Pre-training graph neural networks for cold-start users and items representation. In WSDM’21, pp. 265-273

    2021

  51. [59]

    (2019, July)

    Zhang, J., Hao, B., Chen, B., Li, C., Chen, H., & Sun, J. (2019, July). Hierarchical reinforcement learning for course recommendation in MOOCs. In AAAI’19, V ol. 33, No. 01, pp. 435-442

    2019

  52. [60]

    Chen, J., Wang, F., Pang, S., Chen, M., Xi, M., Zhao, T., & Yin, J. (2025). A Privacy Policy Text Compliance Reasoning Framework with Large Language Models for Healthcare Services. Tsinghua Science and Technology, 30(4), 1831-1845

    2025

  53. [61]

    FT-HashRAG: Combining Hash Retrieval-Augmented Generation with Fine-tuning for Universal Explainable Recommendation

    Zhang Z, Hao B, Wang Y , et al. FT-HashRAG: Combining Hash Retrieval-Augmented Generation with Fine-tuning for Universal Explainable Recommendation. Tsinghua Science and Technology, 2025

    2025

  54. [62]

    (2020, September)

    Hao, B., Zhang, J., Li, C., Chen, H., & Yin, H. (2020, September). Recommending courses in MOOCs for jobs: an auto weak supervi- sion approach. In ECML-PKDD’20, pp. 36-51. Zixuan Zhangis currently an undergraduate in Capital Normal University . His main research interests are ...

    2020

  55. [64]

    These datasets train the model to identify potential conflicts and generate appropriate warning messages, covering the following situations (as shown in Table X)

    Conflict and Absent Knowledge:To address the issue of conflicting knowledge, we develop specialized instruction datasets containing contradictory information scenarios, in- cluding Differences in Medical Theories, Conflicting Informa- tion Sources and Practical Problems. These...

  56. [65]

    Proper Traditional Chinese Medicine practice requires individualized syndrome differentiation and treatment

    Clinical Risks:To clarify the auxiliary role of ZhiFang- DanTai , we have incorporated the following disclaimer to emphasize that this system serves as a decision-support tool rather than a substitute for clinical judgment: ”Important Note: The prescription recommendations pro...

  57. [66]

    directs herbs up- ward

    Si Wu Tang (blood/qiharmonizer):Angelica, Safflower - eases palpitations.2) Ba Zhen Tang (multi-deficiency tonic):Ginseng, Astragalus - boosts energy.3) Astragalus-Codonopsis (qi booster): Improves fatigue.4) Ginseng-Astragalus (spleen strengthener):En- hances stamina Honeysuc...

  58. [2022]

    His research interests include data mining, large language model for recommendation, and large language model for medical diagnosis

    He is currently an assistant professor in Capital Normal University . His research interests include data mining, large language model for recommendation, and large language model for medical diagnosis. He has published 10+ papers in the top Data mining (TOIS, WSDM, AAAI) and ...

    2014

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.