Yuyin Zhou releases ClinSeekAgent, an open-source clinical AI agent that seeks its own evidence
Zhou's team links raw EHR, web, and chest X-ray tools, and reports open-source SOTA after distilling Claude Opus 4.6 into a 35B model.
By Ryan Merket ยท Published
Why it matters
Most clinical AI work evaluates on curated inputs; ClinSeekAgent instead automates evidence seeking across EHRs, web, and imaging and shows reported gains, then distills that behavior into an open 35B model.
Yuyin Zhou (@yuyinzhou_cs) introduced ClinSeekAgent, an automated agentic framework for multimodal evidence seeking in clinical AI, in a thread on X. The push is simple but pointed: stop handing models curated packets of evidence and make them find, weigh, and link the facts themselves.
"Clinical AI shouldn't just consume evidence handed to it - it should actively seek evidence," Zhou wrote on X, arguing that real clinical workflows require deciding where to look, what to retrieve, and how to proceed.
What shipped
ClinSeekAgent exposes a unified tool space designed for long-horizon, plan-act-replan behavior across heterogeneous data sources:
- 11 raw-EHR retrieval tools
- 3 web-search tools
- 6 chest X-ray imaging tools
There is no fixed retrieval order. The agent plans, invokes tools, revises its hypotheses as new information arrives, and integrates the evidence into a clinical answer. The code and models are released openly on GitHub, with a technical report on arXiv.
Why Zhou is building this
Most medical-agent evaluations assume the hard part is done: the relevant evidence is pre-picked and injected into the prompt. Zhou has been pushing back on that assumption in her timeline for months and reiterated it in this launch: "Most medical-agent benchmarks are too idealized - they assume the relevant evidence has already been hand-picked and packaged... Real clinical workflows don't." That critique frames ClinSeekAgent as a testbed for agents that gather and justify their own inputs, not just reason over neatly curated snippets.
Distillation and results
Beyond acting as an inference-time agent, ClinSeekAgent doubles as a training pipeline. Zhou says the team distilled Claude Opus 4.6's search trajectories into Qwen3.5-35B-A3B, yielding the distilled model ClinSeek-35B-A3B. On AgentEHR-Bench, Zhou reports a jump from 22.1 to 34.0 (+11.9), which she describes as open-source SOTA in a follow-up X post.
The arXiv paper also reports inference-time gains when plugging ClinSeekAgent into existing models. On text-only EHR tasks, Claude Opus 4.6 improves from 60.0 to 63.2 overall F1, and MiniMax M2.5 from 43.1 to 47.3. On multimodal chest X-ray tasks, Claude Opus 4.6 improves from 47.5 to 62.6 (+15.1). The authors frame these as evidence that active evidence seeking can lift both proprietary and open models when the environment exposes raw records, web, and imaging tools.
Open, team-led effort
Zhou credited leading contributors Juncheng Wu (@JJwu41867797) and Letian Zhang, and thanked collaborators Yuhan Wang, Haoqin Tu (@HaoqinT), Hardy Chen (@HardyChen266091), Zijun Wang (@zijun_wang2002), and Cihang Xie (@cihangxie). The repository is hosted under the UCSC-VLAA organization path on GitHub. Paper, code, and models are linked from the thread and the repo.
For operators and builders, the takeaway is not just a new benchmark score. ClinSeekAgent packages a working pattern for agentic retrieval across messy clinical surfaces, plus a recipe to distill that behavior into a smaller model. Whether you are prototyping clinical decision support or eyeing other regulated domains with fragmented data, the release offers a concrete starting point with reproducible tooling and evaluations.