DISC-MedLLM: Bridging General Large Language Models and Real-World Medical Consultation (Aug 2023, arXiv)

Forum topic · 小凯 · 2026-07-05

Summary

DISC-MedLLM is an open medical large language model introduced in an August 2023 arXiv paper (arXiv:2308.14346) by Zhijie Bao, Wei Chen, Shengze Xiao, Kuang Ren, Jiaao Wu, Cheng Zhong and colleagues. The work aims to bridge general-purpose LLMs and real-world medical consultation scenarios, targeting multi-turn patient-doctor dialogue, medical information provision, and auxiliary triage. The authors construct a supervised fine-tuning instruction dataset that combines multi-turn conversational data derived from real consultation records, knowledge-base-grounded QA pairs, and patient-agent simulated dialogues, enabling the model to follow instruction-driven, service-oriented behavior. Built on an open-source Chinese LLM base, DISC-MedLLM is evaluated against general and medical LLM baselines, with reported improvements in faithfulness, contextual awareness of multi-turn interactions, and goal-oriented triage behavior. The paper also discusses hallucination mitigation, evaluation methodology, and deployment considerations for trustworthy medical dialogue systems. Model resources were released to the community, making the work a notable early example of adapting general LLMs to vertical-domain healthcare services.

DISC-MedLLM: Bridging General Large Language Models and Real-World Medical Consultation

Paper: arXiv:2308.14346
Authors: Zhijie Bao, Wei Chen, Shengze Xiao, Kuang Ren, Jiaao Wu, Cheng Zhong, et al. (9 authors in total)
Published: August 2023 (arXiv preprint)
Category: Vertical-domain LLMs (Healthcare)

Overview

DISC-MedLLM is a medical-domain large language model designed to bridge the gap between general-purpose LLMs and real-world medical consultation services. Rather than focusing solely on medical exam QA, the project targets practical service scenarios: multi-turn patient-doctor dialogues, medical information provision, and auxiliary triage.

Key Ideas

Service-oriented objective: The model is tuned for instruction-driven, multi-turn medical conversations rather than single-turn benchmark answering, reflecting how patients actually interact with online consultation platforms.
Dataset construction pipeline: The authors build a supervised fine-tuning dataset that combines:
Multi-turn conversational data derived from real online consultation records;
Knowledge-base-grounded QA pairs to inject reliable medical knowledge;
Patient-agent simulated dialogues to expand coverage of triage and dialogue behaviors.
Behavioral goals: faithfulness (reducing hallucinated medical claims), multi-turn contextual awareness, and goal-oriented triage (guiding patients to appropriate departments or care).

Model and Training

The model is initialized from an open-source general-purpose Chinese LLM base and adapted via supervised fine-tuning on the curated medical instruction dataset, following a standard instruction-tuning recipe adapted to the healthcare vertical.

Evaluation

Experiments compare DISC-MedLLM against general-purpose and medical LLM baselines, with both automatic metrics and human/LLM evaluation emphasizing the service-oriented properties listed above. Readers should consult the original PDF tables for exact quantitative results.

Significance

Released in August 2023, DISC-MedLLM was among the early open efforts to adapt general LLMs to trustworthy healthcare dialogue, including discussion of hallucination mitigation, evaluation methodology, and deployment considerations. The model and data resources were released publicly.

References

Original paper: DISC-MedLLM: Bridging General Large Language Models and Real-World Medical Consultation. https://arxiv.org/abs/2308.14346