Autonomous multi-agent research with breadth- and depth-first search and shared evidence graphs for long-horizon discovery, outperforming SOTA agents on HLE and others.
Zifeng Wang
I develop language models and AI agents for reasoning, tool use, and autonomous scientific discovery, combining SFT, RLVR, and agentic reinforcement learning with executable research environments and scalable agent systems. My publications have received 4,500+ citations with h-index of 30+. Selected works include MedCLIP, TransTab, TrialGPT, DeepEvidence, DSWizard, and TrialPanorama.
I received my Ph.D. in Computer Science from UIUC. Previously, I co-founded Keiji.AI and led a team of 10+ researchers and engineers, building AI co-scientists deployed at 10+ biopharma organizations.
Selected work
All publications ↗An 8B model trained with SFT and RL outperformed generic 70B models on eight tasks built from 1.6M trials.
Starting in 2024, pioneered the shift from coding LLMs to biomedical research agents, introducing 293 paper-level executable tasks with evaluation rubrics and achieving 74% accuracy vs. <40% for standalone LLMs.
Automatically synthesized 633K training instructions from 20K+ reviews and 450K+ papers, enabling a 7B model to search and reason over 30M+ biomedical publications, outperforming GPT-4o on multiple tasks.
Starting in 2023, pioneered LLM-driven literature reviews from research questions, automating search, screening, and extraction; a human-guided pilot reduced screening time by 44% and extraction time by 63%.
Harnessed LLMs to automate patient recruitment through clinical trial matching, reducing screening time by 42.6% in a pilot; featured in the NIH Director’s Blog and NIH’s FY2027 Congressional Budget Justification.
Experience
Background ↗2025–2026
Co-founder & Chief AI Scientist · Keiji.AI
Led research and engineering across scientific agents, evaluation, and production deployment, translating research into tools for clinical development.
Recent updates
- Oct 2026Listed in the 2026 Stanford–Elsevier Top 2% Scientists database (single-year impact).
- Jul 2026Our tutorial on LLMs for medical research was accepted in Nature Protocols.
- Jun 2026New work on clinical reasoning, deep evidence research, and data science agents.