Zheng Chen

Research profile

Research summary

Zheng Chen is an early-stage researcher whose catalogued work centers on artificial intelligence for Traditional Chinese Medicine, approached from three complementary angles: clinical language models, prescription generation, and the data infrastructure beneath them. All three papers in Evolve's catalog are 2026 arXiv preprints with Zheng Chen in first-author position — none carries a corresponding-author record — and all were produced within a compact circle of seven distinct co-authors, including Haoxuan Li, Peiwu Qin, and Zhicheng Du. The portfolio splits almost evenly across three fields (~33% each): medical AI/clinical NLP for TCM informatics, medical natural language processing, and database systems/knowledge-graph data management. LingLan builds a multimodal TCM diagnosis LLM with supervised fine-tuning and LoRA, reporting a 103.5% relative accuracy gain to 62.72% — a claim Evolve flags for malformed metrics and inconsistent confusion-matrix counts. A second preprint frames prescription generation as structured reasoning with chain-of-thought distillation and DPO alignment for medication safety; its pipeline shows real effect, though the "7B surpasses GPT-5" headline rests on inconsistent PQS reporting and very small margins. TCMaster contributes confidence-aware querying and workload-guided physical design over a 221K-node multi-source TCM knowledge graph, with honest, modest gains. Reproducibility averages 58 across the three papers (high, medium, and low readiness), while output, impact, and activity metrics remain unscored for lack of citation and publication-volume data.

Research keywords

Highlights

• TCMaster: confidence-aware querying on a 221K-node TCM knowledge graph

  The preprint pairs confidence-aware querying and Cypher rewriting with workload-guided physical design for multi-source TCM knowledge graphs. Evolve's verdict: honest, modest gains, with the 20pp QA win resting on a self-built benchmark — and it carries the highest reproducibility readiness (high) of the three papers.

• LingLan: multimodal TCM diagnosis LLM with tongue and pulse signals

  Built with supervised fine-tuning and LoRA, LingLan reports a 103.5% relative accuracy gain to 62.72% for TCM diagnosis. Evolve flags malformed metrics and inconsistent confusion-matrix counts, and reproducibility readiness is low pending scrutiny.

• Structured reasoning and DPO alignment for TCM prescription generation

  The pipeline combines chain-of-thought distillation and knowledge-driven alignment around syndrome differentiation, synergy, and medication safety. Evolve finds real practical effect, but the "7B surpasses GPT-5" headline rests on inconsistent PQS reporting and very small margins; readiness is medium.

Timeline

1. 2026First-author arXiv preprint TCMaster: confidence-aware querying and workload-guided physical design for multi-source TCM knowledge graphs.
2. 2026First-author arXiv preprint LingLan: a multimodal TCM diagnosis LLM combining tongue and pulse diagnosis.
3. 2026First-author arXiv preprint on structured reasoning and knowledge-driven alignment for TCM prescription generation.

Papers in the Evolve catalog

• 2026-09-22first author

  Syndrome, Synergy, and Safety: Structured Reasoning and Knowledge-Driven Alignment for TCM Prescription Generation

• 2026-09-22first author

  LingLan: An Advancing Traditional Chinese Medicine Diagnosis LLM with Multimodal Data

• 2026-09-22first author

  TCMaster: Confidence-Aware Querying and Workload-Guided Physical Design for Multi-Source Traditional Chinese Medicine Knowledge Graphs

Frequent co-authors