DISC-MedLLM: Bridging General Large Language Models and Real-World Medical Consultation
Paper: https://arxiv.org/abs/2308.14346 (arXiv, August 2023) Authors: Zhijie Bao, Wei Chen, Shengze Xiao, Kuang Ren, Jiaao Wu, Cheng Zhong, et al. (9 authors total) Type: Academic paper Section: Verticals
Overview
DISC-MedLLM is a medical-domain large language model designed to bridge general-purpose LLMs and real-world medical consultation scenarios. Unlike general chatbots, medical consultation requires multi-turn dialogue management, accurate domain knowledge, and communication aligned with patient needs. The paper addresses the question of how to specialize a general LLM for this vertical.
Key points
- Goal: Adapt a general large language model to real-world medical consultation, emphasizing multi-turn interaction, factual reliability, and patient-aligned responses.
- Data strategy: A three-fold data construction approach is described, combining:
- Multi-turn instruction data for conversational consultation behavior;
- Knowledge-infused Q&A pairs to inject medical knowledge;
- Doctor-patient dialogue data reflecting real consultation patterns.
- Positioning: The work belongs to the wave of vertical-domain LLMs (medical, legal, finance) built on instruction tuning of open or general base models.
- RAG and knowledge access: Medical LLMs must ground outputs in reliable knowledge sources to control hallucination and safety risks.
- Evaluation challenges: Offline metrics alone do not capture clinical usefulness; human expert review and safety auditing are necessary.
- Engineering constraints: Latency, cost, interpretability, and safety policies are hard requirements for medical deployment — academic benchmarks alone are insufficient.
- An interpretable ensemble of graph and language models for improving s…
- Applying Deep Learning to Ads Conversion Prediction in Last Mile Deliv…
- Automated Query-Product Relevance Labeling using Large Language Models…
- Behavior Modeling Space Reconstruction for E-Commerce Search, Jan 2025…
- Behavior-driven query similarity prediction based on pre-trained langu…
- Better to Ask in English: Cross-Lingual Evaluation of Large Language M…
- Original paper: DISC-MedLLM: Bridging General Large Language Models and Real-World Medical Consultation. arXiv:2308.14346.
Context within Search / Retrieval research
Although a medical-domain model, DISC-MedLLM sits at the intersection of LLM verticals and retrieval-augmented systems:
Limitations and open questions
As with most vertical LLM papers of this period, potential limitations include benchmark distribution mismatch with real patient populations, unclear cross-lingual generalization, hallucination and safety risks in open-ended consultation, and evaluation subjectivity in multi-turn medical dialogue. Quantitative results should be verified against the original PDF.