Pith. sign in

REVIEW 5 major objections 6 minor 47 references

GAP: Graph-Assisted Prompts for Dialogue-based Medication Recommendation

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

Pith's one-line read Graph-assisted prompts lift LLM medication-recommendation accuracy by encoding the patient's dialogue as an explicit graph, then querying external knowledge along physician-defined paths.

desk verdict GAP is a promising integration of graph-construction over LLM extraction with KG/LLM/internet prompts for medication recommendation, but the unevaluated extraction of pregnancy-related states is a load-bearing gap that blocks acceptance as-is. read the letter →

arxiv 2505.12888 v1 pith:RRGQMYVS submitted 2025-05-19 cs.CL

classification cs.CL
keywords dialogue-basedmedicationrecommendationpatient-centricgraphgraph-assistedpromptingknowledgeretrievalretrieval-augmentedgenerationlargelanguagemodelsmedicaldialoguesystemsdrugcontraindicationsafety
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

GAP sets out to prove that a doctor-side large language model can recommend medications more accurately and safely if it first turns the patient's dialogue history into an explicit graph of medical concepts and their states, and then uses that graph to query external medical knowledge. The paper claims that across multi-turn dialogue, plain LLM prompting loses fine-grained facts such as pregnancy or bronchitis history, so the graph acts as a structured memory that keeps those facts in view. On the dialogue-based medication benchmark, GAP raises Jaccard from 30.15 (a knowledge-graph RAG baseline) to 39.28 and F1 from 42.44 to 54.27, and in simulated diagnostic interviews it scores 0.63 versus 0.51 for a strong general LLM.

What carries the argument

The patient-centric graph $G_p$: a graph whose root is a patient node, with edges to normalized medical concepts (disease, symptom, medication) and directed edges from each concept to its slot-value states (e.g., treatment effective, past medical history). It is constructed from LLM-extracted concepts and states across dialogue turns, and it is then joined with an external medical knowledge graph to form a neighborhood graph $N_{G_p}$. From this structure GAP generates neighborhood prompts (top-1 relation-selected facts from $N_{G_p}$) and path-based prompts (paths matched against pre-defined medical schemas, converted into queries for KG verification, LLM reasoning, and Internet search). The graph and prompts are concatenated with the dialogue history as the final LLM input.

What would settle it

Run the extraction step alone on the DialMed dialogues and compare its concept and state recall against a clinician-verified gold set, especially for states such as pregnancy, allergy history, and past medication use. If extraction recall is substantially below 1.0, or if a dialogue whose gold recommendation changes with the state 'pregnant' produces the same graph with and without that state, then GAP's safety advantage would fail exactly where the paper claims it.

Watch

Extended reading notes

Core claim

The central claim is that dialogue-based medication recommendation improves when the LLM is prompted not only with the raw dialogue but with a patient-centric graph built from extracted medical concepts (diseases, symptoms, medications) and their attributed states (pregnancy, duration, past medical history), plus two kinds of derived prompts: neighborhood prompts that pull nearby facts from an external medical knowledge graph, and path-based prompts that query the knowledge graph, the LLM itself, and the Internet along physician-defined schemas. GAP reports that this combination outperforms existing LLM-based recommendation methods on the DialMed benchmark with Jaccard 39.28 and F1 54.27, and improves diagnostic interviewing information density (0.19) and score (0.63). The paper presents this as evidence that explicit graph memory plus multi-source retrieval can reduce non-factual and contraindicated medication responses.

Load-bearing premise

The load-bearing premise is that the LLM extraction step catches all medically relevant concepts and states in the dialogue, since the paper never measures extraction accuracy and a missed state like pregnancy drops out of the graph entirely.

Editorial extensions

If this is right

  • If GAP's gains hold, any black-box LLM can be turned into a more reliable recommender by adding a dialogue-derived graph plus retrieval prompts, with no retraining or parameter updates.
  • Ablation results show that both prompt types matter: removing neighborhood prompts drops F1 by about 19%, and removing path-based prompts drops F1 by about 11%, so the retrieval design carries much of the benefit.
  • The path schema is modular: the paper notes it can be re-targeted from acquiring treatments to suitable food, medical tests, or lifestyle advice, so the same architecture extends to other dialogue-based health services.
  • Because GAP is orthogonal to training-time methods, it can be stacked on top of fine-tuned medical LLMs and chain-of-thought prompting, potentially compounding their gains.

Reading between the lines

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

  • The safety claim rests entirely on the LLM extraction step: the paper never measures extraction accuracy, so if a state like 'pregnant' is missed, the graph omits it and a contraindicated drug can still be recommended; a testable extension is to measure concept and state extraction recall separately and to inject a global state summary from an independent extractor.
  • The graph is a lossy textual summary of the dialogue; replacing the linearized triple string with a graph neural network, or storing uncertainty about extracted states, could improve robustness to extraction noise.
  • Because results are reported only on a Chinese dialogue dataset with a specific set of Chinese medical knowledge sources, the framework's portability to English sources and other knowledge graphs is open; a replication with English KGs would test whether the benefit is language- or KG-dependent.
  • The diagnostic-interviewing gain is a single aggregate score; decomposing it into aspects and information catches would show whether GAP's improvement comes from asking more complete questions or from eliciting more key items from the simulated patient.
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 GAP, a framework for dialogue-based medication recommendation. GAP extracts medical concepts and their states from a dialogue using LLM-based two-stage extraction, builds a patient-centric graph, and generates neighborhood prompts and path-based prompts by querying external knowledge sources (CMeKG, Disease-KB, LLM reasoning, and Internet search). The framework is evaluated on the DialMed medication recommendation dataset and on a diagnostic interviewing task using an LLM-as-Patients simulator, reporting Jaccard/F1 improvements over LLM baselines and a diagnostic score of 0.63. The paper includes ablations showing the contribution of each prompt type and an error analysis.

Significance. If the results hold, GAP provides a practical plug-and-play method for grounding LLM-based medication recommendations in an explicit patient-specific representation, with consistent gains over KG-RAG and few-shot CoT prompting on DialMed, and it demonstrates a promising direction for integrating structured dialogue memory with heterogeneous knowledge sources. The ablation study is well-designed and shows clear contributions from both neighborhood and path-based prompts. The main contributions are the patient-centric graph construction and the multi-source prompt generation. However, the central mechanism depends on upstream extraction components that are not specified or evaluated, and the reported configuration contains a direct inconsistency about the underlying LLM, so the current evidence is insufficient to support the paper's safety-related claims.

major comments (5)
  1. [§3.2, Appendix A.1, Table 5] The extraction equations (1)-(2) and the prompts in Tables 4-5 cover medical concepts and their states, but the patient-characteristic extraction (gender, age, blood pressure, pregnancy) that Appendix A.1 lists as part of the patient-centric graph is never specified with a prompt, an equation, or an evaluation. The motivating safety example in Figure 1 and the path-based prompts in Figure 3 depend on detecting pregnancy correctly; a missed pregnancy directly contaminates the graph, neighborhood prompts, and path-based prompts. The paper reports no extraction precision, recall, or F1 on DialMed or any corpus, so the mechanism claimed to reduce unsafe recommendations is unverified.
  2. [§4.1, §4.2, Appendix A.3] The foundation LLM is inconsistent. Section 4.1 states 'We use DeepSeek-V3 [4] as the foundation LLM' and Section 4.2 says DeepSeek-V3 is utilized for extraction and prompt generation, while Appendix A.3 states 'We use ChatGPT-3.5-Turbo (version 1106) for the experiments.' This is a direct contradiction that affects the reproducibility and interpretation of all reported results; the authors must clarify which model was used for which component and reconcile the two statements.
  3. [Tables 1 and 3] The paper reports means and standard deviations over five runs but no statistical significance tests (e.g., paired bootstrap or t-test). The central claim that 'GAP outperforms other LLM-based baselines on all department cases' (Section 4.3) requires statistical validation, especially for comparisons where the gaps are smaller (e.g., Respiratory Jaccard: GAP 29.33 vs. KG-RAG 27.48). Please report significance tests or confidence intervals.
  4. [Abstract, §4.3] The abstract claims that GAP 'retrieving information from multiple sources to reduce the non-factual responses.' However, the evaluation only measures medication-recommendation accuracy (Jaccard/F1) and a diagnostic score; it does not measure factual correctness or hallucination in the generated text. The error analysis (Figure 4) is conducted on CoT errors, not on GAP errors. Thus, the specific benefit of 'non-factual' reduction is not demonstrated; please provide direct evidence or temper the claim.
  5. [§3.3] The linking of extracted mentions to KG entities is described as 'simple linking methods (e.g., edit distance, synonym list)' but its accuracy is never measured. Since neighborhood prompts are retrieved via this linking, incorrect normalization will propagate to the retrieved knowledge. Please evaluate the linking component or discuss its failure modes in the context of the safety claims.
minor comments (6)
  1. [§4.2] The sentence 'The context sliding window size of extraction is set to 4 (i.e., k = 1 in Equation 2), and set to ∞ in diagnostic task and the recommendation task, respectively' is ambiguous; please clarify which value applies to which task and how the window size relates to k.
  2. [§4.1, Table 2] The diagnostic evaluation relies on the authors' own LLM-as-Patients framework [16] and LLM-based scoring; please discuss the potential for self-evaluation bias and provide details on the agreement of the five LLM evaluators.
  3. [§4.3] The error analysis in Figure 4 is conducted on CoT errors; an error analysis of GAP itself would help identify remaining failure modes for this framework.
  4. [Table 7] The prompt in Table 7 uses the term 'Respirator' for the department; this should be 'Respiratory'.
  5. [§3.4] The hyperparameters k1, k2, and the context window size are set heuristically; a sensitivity analysis would strengthen the robustness of the reported results.
  6. [Appendix A.2] The path schema S is said to be defined by medical professionals, but no details about the panel, their instructions, or inter-annotator agreement are provided; please add these details.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-citation in the diagnostic-interviewing evaluation; the core DialMed medication-recommendation result is externally benchmarked and not circular.

  1. self citation load bearing [Section 4.1, Datasets / Evaluation metrics (diagnostic task), and Table 2]
    "LLM as Patients [16] is a novel method to evaluate diagnostic interviewing abilities of MDS... For the diagnostic task, we follow the setting of [16] by adopting the weighted coverage rate of Aspects and Information in the examination checklist to obtain the score of diagnostic process... the scores come from the ensemble of 5 different LLM evaluators."

    Reference [16] (Li, Zeng, Zhong, Zhang, Zhang, Zou) is the authors' own prior work, with overlap (Zhong, Li, Zou) with the present paper. The claimed diagnostic improvement (GAP 0.63 vs ChatGPT 0.51) is measured entirely within that self-defined LLM-as-Patients protocol and by LLM evaluators, so the secondary diagnostic claim rests on a self-citation rather than on an external benchmark. This is not a construction-level circularity (no fitted value is reused as the target), and it does not bear on the DialMed medication-recommendation result, which uses an external dataset and ground-truth medication sets.

full rationale

The central DialMed derivation is self-contained: GAP extracts concepts and slot-value states with LLM calls (Eqs. 1-2), builds a patient-centric graph, adds KG/LLM/Internet knowledge via neighborhood and path prompts, and generates a recommendation (Eq. 6). No parameter is fitted to DialMed labels and the recommendation targets are not fed into the extraction or prompt modules, so the Jaccard/F1 gains are not forced by construction. The ablation (Table 3) shows the prompts contribute beyond the graph, and the comparison is against an external benchmark. The only circularity-adjacent element is the diagnostic-interviewing experiment: it uses the authors' own LLM-as-Patients framework and LLM evaluators from [16], which is a self-citation and self-evaluation for that secondary claim. Because that claim is exploratory and not the central benchmark, I rate this as a minor self-citation issue (2), not a partial or full circularity. The inconsistent LLM identity (DeepSeek-V3 in Sec. 4.2 vs ChatGPT-3.5-Turbo in App. A.3) is a correctness/reproducibility concern, not a circularity.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on an empirical pipeline rather than a formal derivation. No formal proof or machine-checked artifact is provided. The main assumptions are that LLM extraction is reliable, external knowledge sources are correct and current, the hand-crafted path schema is clinically adequate, and the closed candidate list is a fair test setting.

free parameters (6)
  • context window size k = k = 1 (window 4) / infinity
    Sets how many surrounding utterances are used to judge concept states in Eq. (2); chosen by hand and affects extraction quality.
  • max neighborhood prompts k1 = 3
    Number of facts selected from the KG neighborhood in Eq. (3); chosen by hand.
  • max path-based prompts k2 = 3
    Number of matched paths used for retrieval in Eq. (4); chosen by hand.
  • LLM temperature = 0.2
    Sampling temperature for all LLM calls; chosen by hand for stability.
  • ICL demonstrations per stage = 3
    Number of few-shot demonstrations used for extraction, state identification, and reasoning; chosen by hand.
  • evaluation runs averaged = 5
    Table 1 averages five runs; affects reported variance but is not fitted to the target.
assumptions (6)
  • domain assumption The LLM used for extraction returns correct medical concept mentions and slot-value states in Eqs. (1)-(2).
    No extraction-quality evaluation is provided; if extraction misses a state (e.g., pregnancy), the graph and prompts inherit the error.
  • domain assumption External medical KGs (CMeKG, Disease-KB) and supplementary knowledge contain correct and current contraindication and treatment facts.
    KG verification and path prompts rely on these sources; outdated or incomplete KGs would undercut safety claims.
  • ad hoc to paper The pre-defined medical path schema S (Appendix A.2) captures the clinically relevant decision paths for medication selection.
    The schema is crafted for this paper and not empirically validated; different schemas would change prompts and results.
  • domain assumption Internet search in Section 3.4 returns relevant and accurate medical information for generated queries.
    The method depends on open-ended search results; search quality and engine are unspecified, making results environment-dependent.
  • domain assumption Providing the entire candidate medication list to the LLM is a valid proxy for dialogue-based recommendation.
    Section 4.2 states the full candidate list is given; in real consultations no such closed list exists, so deployment transfer is untested.
  • domain assumption The LLM-as-Patients diagnostic evaluation from [16] is a valid measure of diagnostic ability.
    Used for the diagnostic task; the framework and evaluation come from prior work by overlapping authors, so self-assessment risk is present.
invented entities (1)
  • Patient-centric dialogue graph G_p
    purpose: Represents extracted medical concepts, states, and patient attributes from dialogue and connects to external KG entities.
    It is an internal representational device introduced by GAP; no external falsifiable handle exists. Its value is observed only through end-task metrics, and extraction errors propagate directly into it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GAP: Graph-Assisted Prompts for Dialogue-based Medication Recommendation." pith.science (2026). https://pith.science/paper/RRGQMYVS

@misc{pith2026250512888,
  author       = {Pith},
  title        = {Pith review of: GAP: Graph-Assisted Prompts for Dialogue-based Medication Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRGQMYVS}},
  note         = {Machine review of arXiv:2505.12888}
}
read the original abstract

Medication recommendations have become an important task in the healthcare domain, especially in measuring the accuracy and safety of medical dialogue systems (MDS). Different from the recommendation task based on electronic health records (EHRs), dialogue-based medication recommendations require research on the interaction details between patients and doctors, which is crucial but may not exist in EHRs. Recent advancements in large language models (LLM) have extended the medical dialogue domain. These LLMs can interpret patients' intent and provide medical suggestions including medication recommendations, but some challenges are still worth attention. During a multi-turn dialogue, LLMs may ignore the fine-grained medical information or connections across the dialogue turns, which is vital for providing accurate suggestions. Besides, LLMs may generate non-factual responses when there is a lack of domain-specific knowledge, which is more risky in the medical domain. To address these challenges, we propose a \textbf{G}raph-\textbf{A}ssisted \textbf{P}rompts (\textbf{GAP}) framework for dialogue-based medication recommendation. It extracts medical concepts and corresponding states from dialogue to construct an explicitly patient-centric graph, which can describe the neglected but important information. Further, combined with external medical knowledge graphs, GAP can generate abundant queries and prompts, thus retrieving information from multiple sources to reduce the non-factual responses. We evaluate GAP on a dialogue-based medication recommendation dataset and further explore its potential in a more difficult scenario, dynamically diagnostic interviewing. Extensive experiments demonstrate its competitive performance when compared with strong baselines.

Figures

Figures reproduced from arXiv: 2505.12888 by the authors.

Figure 1
Figure 1. A medical consultation between a human patient and LLMs. Red, blue, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall diagram of GAP. 3.3 Graph Construction The extracted medical information can be used to build up a graph as the med￾ical dialogue memory, which has been proven to benefit clinical downstream applications [6]. We call it a patient-centric graph (denoted as Gp) as it contains explicit patient nodes and associated medical concept nodes, designed to accom￾modate multiple patients probably mentioned in the whole … view at source ↗
Figure 3
Figure 3. Medication recommendation cases on DialMed. GAP is compared with [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Error distribution on the samples of DialMed. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Prominent schema used for constructing path-based prompts. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 22 canonical work pages

  1. [4]

    DeepSeek-AI: Deepseek-v3 technical report (2025)

  2. [1]

    Baidu: Ernie bot: Baidu’s knowledge-enhanced large language model built on full ai stack technology (2023)

  3. [2]

    CoRR abs/2308.14346 (2023)

    Bao, Z., Chen, W., Xiao, S., Ren, K., Wu, J., Zhong, C., Peng, J., Huang, X., Wei, Z.: Disc-medllm: Bridging general large language mod- els and real-world medical consultation. CoRR abs/2308.14346 (2023). https://doi.org/10.48550/ARXIV.2308.14346

  4. [3]

    CoRR abs/2310.15896 (2023)

    Chen, Y., Wang, Z., Xing, X., Zheng, H., Xu, Z., Fang, K., Wang, J., Li, S., Wu, J., Liu, Q., Xu, X.: Bianque: Balancing the questioning and suggestion ability of health llms with multi-turn health conversations polished by chatgpt. CoRR abs/2310.15896 (2023). https://doi.org/10.48550/ARXIV.2310.15896

  5. [5]

    In: Findings of EMNLP

    Dou, C., Jin, Z., Jiao, W., Zhao, H., Zhao, Y., Tao, Z.: Plugmed: Improving speci- ficity in patient-centered medical dialogue generation using in-context learning. In: Findings of EMNLP. pp. 5050–5066 (2023)

  6. [6]

    In: Findings of EMNLP

    Gao, L., Zhang, X., Wu, X., Ge, S., Zheng, Y.: Dialogue medical information extraction with medical-item graph and dialogue-status en- riched representation. In: Findings of EMNLP. pp. 13311–13321 (2023). https://doi.org/10.18653/V1/2023.FINDINGS-EMNLP.888

  7. [7]

    CoRR abs/2312.10997 (2023)

    Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Guo, Q., Wang, M., Wang, H.: Retrieval-augmented genera- tion for large language models: A survey. CoRR abs/2312.10997 (2023). https://doi.org/10.48550/ARXIV.2312.10997

  8. [8]

    Guo, Q., Cao, S., Yi, Z.: A medical question answering system using large lan- guage models and knowledge graphs. Int. J. Intell. Syst.37(11), 8548–8564 (2022). https://doi.org/10.1002/INT.22955

Show all 47 references
  1. [9]

    CoRR abs/2310.05694 (2023)

    He, K., Mao, R., Lin, Q., Ruan, Y., Lan, X., Feng, M., Cambria, E.: A survey of large language models for healthcare: from data, technology, and applications to accountability and ethics. CoRR abs/2310.05694 (2023). https://doi.org/10.48550/ARXIV.2310.05694

  2. [10]

    In: ICCL

    He, Z., Han, Y., Ouyang, Z., Gao, W., Chen, H., Xu, G., Wu, J.: DialMed: A dataset for dialogue-based medication recommendation. In: ICCL. pp. 721–733. Gyeongju, Republic of Korea (2022)

  3. [11]

    CoRR abs/2401.04088 (2024)

    Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., de Las Casas, D., Hanna, E.B., Bressand, F., Lengyel, G., Bour, G., Lample, G., Lavaud, L.R., Saulnier, L., Lachaux, M., Stock, P., Subrama- nian, S., Yang, S., Antoniak, S., Scao, T.L...

  4. [12]

    CoRRabs/2312.15883 (2023)

    Jiang, X., Zhang, R., Xu, Y., Qiu, R., Fang, Y., Wang, Z., Tang, J., Ding, H., Chu, X., Zhao, J., Wang, Y.: Think and retrieval: A hypothesis knowledge graph enhanced medical large language models. CoRRabs/2312.15883 (2023). https://doi.org/10.48550/ARXIV.2312.15883

  5. [13]

    In: ICML

    Kandpal, N., Deng, H., Roberts, A., Wallace, E., Raffel, C.: Large language models struggle to learn long-tail knowledge. In: ICML. Proceedings of Machine Learning Research, vol. 202, pp. 15696–15707 (2023)

  6. [14]

    In: SIGKDD

    Le, H., Tran, T., Venkatesh, S.: Dual memory neural com- puter for asynchronous two-view sequential learning. In: SIGKDD. pp. 1637–1645 (2018). https://doi.org/10.1145/3219819.3219981, https://doi.org/10.1145/3219819.3219981 Graph-Assisted Prompts for Dialogue-based Medication...

  7. [15]

    In: NeurIPS (2020)

    Lewis, P.S.H., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive NLP tasks. In: NeurIPS (2020)

  8. [16]

    CoRRabs/2404.13066 (2024)

    Li, Y., Zeng, C., Zhong, J., Zhang, R., Zhang, M., Zou, L.: Leveraging large lan- guage model as simulated patients for clinical education. CoRRabs/2404.13066 (2024)

  9. [17]

    CoRR abs/2402.02803 (2024)

    Liu, Q., Wu, X., Zhao, X., Zhu, Y., Zhang, Z., Tian, F., Zheng, Y.: Large language model distilling medication recommendation model. CoRR abs/2402.02803 (2024). https://doi.org/10.48550/ARXIV.2402.02803, https://doi.org/10.48550/arXiv.2402.02803

  10. [18]

    Liu, S., Wang, X., Zhao, X., Chen, H.: Dkinet: Medication recommendation via domain knowledge informed deep learning (2023)

  11. [19]

    Luo, J., Ye, M., Xiao, C., Ma, F.: Hitanet: Hierarchical time-aware attention net- works for risk prediction on electronic health records. In: KDD. pp. 647–656 (2020)

  12. [20]

    In: Findings of EMNLP

    Mao, K., Dou, Z., Mo, F., Hou, J., Chen, H., Qian, H.: Large language models know your contextual search intent: A prompting framework for conversational search. In: Findings of EMNLP. pp. 1211–1225 (2023)

  13. [21]

    OpenAI: Openai: Introducing chatgpt (2022), https://openai.com/blog/chatgpt

  14. [22]

    In: Findings of EMNLP

    Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N.A., Lewis, M.: Measuring and narrowing the compositionality gap in language models. In: Findings of EMNLP. pp. 5687–5711 (2023). https://doi.org/10.18653/V1/2023.FINDINGS-EMNLP.378

  15. [23]

    In: AAAI

    Shang, J., Xiao, C., Ma, T., Li, H., Sun, J.: Gamenet: Graph aug- mented memory networks for recommending medication combination. In: AAAI. pp. 1126–1133 (2019). https://doi.org/10.1609/AAAI.V33I01.33011126, https://doi.org/10.1609/aaai.v33i01.33011126

  16. [24]

    In: AAAI

    Shen, W., Chen, J., Quan, X., Xie, Z.: Dialogxl: All-in-one xlnet for multi-party conversation emotion recognition. In: AAAI. pp. 13789–13797 (2021)

  17. [25]

    In: AAAI

    Shi, X., Hu, H., Che, W., Sun, Z., Liu, T., Huang, J.: Understanding medical con- versations with scattered keyword attention and weak supervision from responses. In: AAAI. pp. 8838–8845 (2020)

  18. [26]

    CoRR abs/2405.10630 (2024)

    Shi, X., Liu, Z., Du, L., Wang, Y., Wang, H., Guo, Y., Ruan, T., Xu, J., Zhang, S.: Medical dialogue: A survey of categories, methods, evaluation and challenges. CoRR abs/2405.10630 (2024)

  19. [27]

    CoRR abs/2311.17330 (2023)

    Soman, K., Rose, P.W., Morris, J.H., Akbas, R.E., Smith, B., Peetoom, B., Villouta-Reyes, C., Cerono, G., Shi, Y., Rizk-Jackson, A., Israni, S., Nel- son, C.A., Huang, S., Baranzini, S.E.: Biomedical knowledge graph-enhanced prompt generation for large language models. CoRR ab...

  20. [28]

    CoRR abs/2401.01313 (2024)

    Tonmoy, S.M.T.I., Zaman, S.M.M., Jain, V., Rani, A., Rawte, V., Chadha, A., Das, A.: A comprehensive survey of hallucination mitiga- tion techniques in large language models. CoRR abs/2401.01313 (2024). https://doi.org/10.48550/ARXIV.2401.01313

  21. [29]

    Valizadeh, M., Parde, N.: The AI doctor is in: A survey of task-oriented dia- logue systems for healthcare applications. In: ACL. pp. 6638–6660. Dublin, Ireland (2022). https://doi.org/10.18653/v1/2022.acl-long.458

  22. [30]

    CoRR abs/2306.09968 (2023)

    Wang, G., Yang, G., Du, Z., Fan, L., Li, X.: Clinicalgpt: Large language mod- els finetuned with diverse medical data and comprehensive evaluation. CoRR abs/2306.09968 (2023). https://doi.org/10.48550/ARXIV.2306.09968

  23. [31]

    CoRR abs/2302.07257 (2023)

    Wang, S., Zhao, Z., Ouyang, X., Wang, Q., Shen, D.: Chatcad: Interactive computer-aided diagnosis on medical image using large language models. CoRR abs/2302.07257 (2023). https://doi.org/10.48550/ARXIV.2302.07257 16 Zhong et al

  24. [32]

    In: NeurIPS (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E.H., Le, Q.V., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models. In: NeurIPS (2022)

  25. [33]

    CoRR abs/2302.10205 (2023)

    Wei, X., Cui, X., Cheng, N., Wang, X., Zhang, X., Huang, S., Xie, P., Xu, J., Chen, Y., Zhang, M., Jiang, Y., Han, W.: Zero-shot informa- tion extraction via chatting with chatgpt. CoRR abs/2302.10205 (2023). https://doi.org/10.48550/ARXIV.2302.10205

  26. [34]

    CoRR abs/2304.01097 (2023)

    Xiong, H., Wang, S., Zhu, Y., Zhao, Z., Liu, Y., Huang, L., Wang, Q., Shen, D.: Doctorglm: Fine-tuning your chinese doctor is not a herculean task. CoRR abs/2304.01097 (2023). https://doi.org/10.48550/ARXIV.2304.01097

  27. [35]

    In: EMNLP

    Xu, C., Guo, D., Duan, N., McAuley, J.J.: Baize: An open-source chat model with parameter-efficient tuning on self-chat data. In: EMNLP. pp. 6268–6278 (2023)

  28. [36]

    CoRR abs/2310.04408 (2023)

    Xu, F., Shi, W., Choi, E.: RECOMP: improving retrieval-augmented lms with compression and selective augmentation. CoRR abs/2310.04408 (2023). https://doi.org/10.48550/ARXIV.2310.04408

  29. [37]

    In: Findings of ACL

    Xu, K., Hou, W., Cheng, Y., Wang, J., Li, W.: Medical dialogue gener- ation via dual flow modeling. In: Findings of ACL. pp. 6771–6784 (2023). https://doi.org/10.18653/V1/2023.FINDINGS-ACL.423

  30. [38]

    In: CIKM

    Ye, M., Luo, J., Xiao, C., Ma, F.: LSAN: modeling long-term dependencies and short-term correlations with hierarchical attention for risk prediction. In: CIKM. pp. 1753–1762 (2020). https://doi.org/10.1145/3340531.3411864

  31. [39]

    CoRR abs/2310.01558 (2023)

    Yoran, O., Wolfson, T., Ram, O., Berant, J.: Making retrieval-augmented lan- guage models robust to irrelevant context. CoRR abs/2310.01558 (2023). https://doi.org/10.48550/ARXIV.2310.01558

  32. [40]

    CoRRabs/2303.01229 (2023)

    Zakka, C., Chaurasia, A., Shad, R., Hiesinger, W.: Almanac: Knowledge- grounded language models for clinical medicine. CoRRabs/2303.01229 (2023). https://doi.org/10.48550/ARXIV.2303.01229

  33. [41]

    In: EMNLP

    Zeng, G., Yang, W., Ju, Z., Yang, Y., Wang, S., Zhang, R., Zhou, M., Zeng, J., Dong, X., Zhang, R., Fang, H., Zhu, P., Chen, S., Xie, P.: MedDialog: Large-scale medical dialogue datasets. In: EMNLP. pp. 9241–9250. Online (2020). https://doi.org/10.18653/v1/2020.emnlp-main.743

  34. [42]

    In: Findings of EMNLP

    Zhang, H., Chen, J., Jiang, F., Yu, F., Chen, Z., Chen, G., Li, J., Wu, X., Zhang, Z., Xiao, Q., Wan, X., Wang, B., Li, H.: Huatuogpt, towards taming language model to be a doctor. In: Findings of EMNLP. pp. 10859–10885 (2023)

  35. [43]

    In: SIGKDD

    Zhang, Y., Chen, R., Tang, J., Stewart, W.F., Sun, J.: LEAP: learning to prescribe effective and safe treatment combinations for multimorbidity. In: SIGKDD. pp. 1315–1324 (2017). https://doi.org/10.1145/3097983.3098109, https://doi.org/10.1145/3097983.3098109

  36. [44]

    In: HPCC

    Zhao, B., Jiang, Z., Zhang, J., Ma, F., Li, J.: Medical dialogue generation via extracting heterogenous information. In: HPCC. pp. 194–201 (2022). https://doi.org/10.1109/HPCC-DSS-SMARTCITY- DEPENDSYS57074.2022.00058

  37. [45]

    Zhao, Y., Li, Y., Wu, Y., Hu, B., Chen, Q., Wang, X., Ding, Y., Zhang, M.: Medical dialogue response generation with pivotal information recalling. In: KDD. pp. 4763–4771 (2022). https://doi.org/10.1145/3534678.3542674

  38. [46]

    CoRR abs/2310.06117 (2023)

    Zheng, H.S., Mishra, S., Chen, X., Cheng, H., Chi, E.H., Le, Q.V., Zhou, D.: Take a step back: Evoking reasoning via abstraction in large language models. CoRR abs/2310.06117 (2023). https://doi.org/10.48550/ARXIV.2310.06117

  39. [47]

    patient claims positive

    Zhu, Y., Moniz, J.R.A., Bhargava, S., Lu, J., Piraviperumal, D., Li, S., Zhang, Y., Yu, H., Tseng, B.: Can large language models understand context? In: Findings of EACL. pp. 2004–2018 (2024) Graph-Assisted Prompts for Dialogue-based Medication Recommendation 17 A Appendix A.1...

Pith tools

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