Gaoyuan Du
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Gaoyuan Du has a total of 25 co-authors. Here are the top co-authors by number of shared publications:
- Dimitrios Dimitriadis
Amazon
2 shared publications - Xueping Li
Department of Industrial and Systems Engineering, University of Tennessee, Knoxville, USA
2 shared publications - Aaron Chan
Amazon, Bellevue, WA, USA
1 shared publication - Arghya Datta
Amazon, Seattle, WA, USA
1 shared publication - Arushi Gupta
Amazon, Seattle, WA, USA
1 shared publication - Esther Xu Fei
Department of Environmental Health and Engineering, Johns Hopkins University, Baltimore, MD, USA
1 shared publication
Latest Publications
- 29 August 2026
- Published by Elsevier in Journal of Biomedical Informatics
- Vol. 182
Evaluating teacher-supervised small medical language models for institution-constrained clinical QA
- 29 August 2026
- Published by Elsevier in Journal of Biomedical Informatics
- Vol. 182
Evaluating teacher-supervised small medical language models for institution-constrained clinical QA
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9604-9615
SOP-Bench: Complex Industrial SOPs for Evaluating LLM Agents
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 9604-9615
Open AccessSOP-Bench: Complex Industrial SOPs for Evaluating LLM Agents
- Open Access
- 11 June 2026
- Published by Elsevier in SSRN Electronic Journal
Evaluating Teacher-Supervised Small Medical Language Models for Institution-Constrained Clinical QA
Objective: To evaluate whether small language models adapted on public medical benchmarks transfer reliably to institution-constrained open-answer clinical QA, and to identify which adaptation choices—supervision format, optimization objective, and backbone—determine local answer quality and evidence coverage.Methods: We used DistilGPT2 (82M) as the primary student and Llama3 70B as the teacher. We compared four strategies: public-benchmark answer-only adaptation, public-benchmark chain-of-thought adaptation, in-domain teacher-supervised question-answer fine-tuning (QAFT), and direct preference optimization (DPO). We evaluated each strategy on cleaned public benchmarks and on an internal EHR-grounded open-answer task using token-F1, exact match, hallucinated evidence rate, and evidence coverage. We tested robustness via multi-seed reruns, a controlled DPO pair-construction ablation, and a TinyLlama 1.1B cross-backbone replication.Results: Public-benchmark transfer did not produce stable gains under cleaned repeated-seed evaluation. In-domain QAFT consistently outperformed public-benchmark transfer on the local task (F1 0.1310 vs 0.1090). DPO did not improve answer quality uniformly; instead, it shifted models toward shorter, stricter responses with lower evidence coverage. This shift depended on how the rejected response was constructed, not on the DPO objective itself. TinyLlama replication confirmed that in-domain QAFT generalized across backbones, while DPO effects were backbone-specific.Conclusion: For institution-constrained clinical QA, in-domain teacher-supervised fine-tuning is the most reliable adaptation path. Public-benchmark transfer is not a dependable proxy for local utility, and DPO should be interpreted as operating-point control with local validation before deployment. We propose a local evaluation protocol: adapt on the target task, validate DPO as an operating-point control, and interpret groundedness jointly with evidence coverage.
- 11 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEvaluating Teacher-Supervised Small Medical Language Models for Institution-Constrained Clinical QA
Objective: To evaluate whether small language models adapted on public medical benchmarks transfer reliably to institution-constrained open-answer clinical QA, and to identify which adaptation choices—supervision format, optimization objective, and backbone—determine local answer quality and evidence coverage.Methods: We used DistilGPT2 (82M) as the primary student and Llama3 70B as the teacher. We compared four strategies: public-benchmark answer-only adaptation, public-benchmark chain-of-thought adaptation, in-domain teacher-supervised question-answer fine-tuning (QAFT), and direct preference optimization (DPO). We evaluated each strategy on cleaned public benchmarks and on an internal EHR-grounded open-answer task using token-F1, exact match, hallucinated evidence rate, and evidence coverage. We tested robustness via multi-seed reruns, a controlled DPO pair-construction ablation, and a TinyLlama 1.1B cross-backbone replication.Results: Public-benchmark transfer did not produce stable gains under cleaned repeated-seed evaluation. In-domain QAFT consistently outperformed public-benchmark transfer on the local task (F1 0.1310 vs 0.1090). DPO did not improve answer quality uniformly; instead, it shifted models toward shorter, stricter responses with lower evidence coverage. This shift depended on how the rejected response was constructed, not on the DPO objective itself. TinyLlama replication confirmed that in-domain QAFT generalized across backbones, while DPO effects were backbone-specific.Conclusion: For institution-constrained clinical QA, in-domain teacher-supervised fine-tuning is the most reliable adaptation path. Public-benchmark transfer is not a dependable proxy for local utility, and DPO should be interpreted as operating-point control with local validation before deployment. We propose a local evaluation protocol: adapt on the target task, validate DPO as an operating-point control, and interpret groundedness jointly with evidence coverage.