Zijun Yao

Research profile

Research summary

Zijun Yao is a machine learning researcher at the University of Kansas whose work applies reinforcement learning and large language models to electronic health record (EHR) mining. OpenAlex bibliometric data (identity resolved via OpenAlex) records an h-index of 10 and i10-index of 10, with 36 works in the last three years and a steep activity ramp: single-digit annual output in 2017-2019, 7 works in 2021, 9 in 2023, 11 in 2025, and 18 already recorded in 2026. The topic profile is led by topic modeling (41 works), followed by natural language processing techniques (19), multimodal machine learning applications (12), explainable AI (11), and information retrieval and search behavior (7). The Evolve-reviewed catalog reflects the current core: as last author on REFINE (frozen-LLM soft prompts over RL-budgeted patient-specific text-attributed graphs on MIMIC) and Import What You Need (RL-guided per-visit knowledge-graph import on MIMIC-III/IV), and as a middle author on an analysis showing RLVR gains are largely recoverable by budgeted base-model search. Collaboration scores 44/100 with 8 distinct recent coauthors; reproducibility scores 55/100 from one evaluated paper.

Research keywords

Highlights

• REFINE: RL-budgeted graph context with frozen-LLM soft prompts for EHR codes

  Last-authored 2026 arXiv paper builds patient-specific text-attributed graph context selected under a reinforcement-learning budget and pairs it with frozen-LLM soft prompts for medical concept representation; MIMIC gains look solid, though Evolve flags missing fold variance and significance tests (reproducibility: medium).

• Import What You Need: RL-guided per-visit knowledge-graph augmentation for EHR

  Last-authored 2026 arXiv work learns when and how to import external knowledge into EHR graphs, choosing Soft, Hard, or Skip per visit with reinforcement learning; gains over static augmentation on MIMIC-III/IV are modest but backed by careful controls and audits, with high reproducibility readiness.

• Budgeted base-model search recovers much of RLVR's reasoning gains

  Middle-authored 2026 arXiv analysis argues that RL with verifiable rewards (RLVR) gains are largely recoverable by budgeted base-model search (BOPTR); a rigorous and honest study, though the edge over a greedy baseline falls within noise (high reproducibility readiness).

Timeline

1. 2017First recorded works appear in the OpenAlex publication record (1 publication).
2. 2021Annual output rises to 7 publications, starting a sustained climb.
3. 20239 publications recorded as the NLP and topic-modeling core consolidates.
4. 202511 publications recorded, the strongest full year to date.
5. 202618 works recorded so far; three arXiv preprints enter the Evolve catalog — two as last author on knowledge-graph-augmented EHR prediction, one middle-authored on RLVR versus budgeted search.

Papers in the Evolve catalog

• 2026-09-03

  REFINE: LLM Refinement over Budgeted Text-Attributed Graphs for Personalized Medical Concept Representation

• 2026-09-01

  Import What You Need: Learning When and How to Augment EHR Graphs with External Knowledge

• 2026-09-01

  From Base Rollouts to RL Reasoning: A Budgeted Search Perspective

Frequent co-authors