UCSC-VLAA/ClinSeek-35B-A3B
ClinSeek-35B-A3B is a 35.1 billion parameter language model developed by UCSC-VLAA, fine-tuned from Qwen/Qwen3.5-35B-A3B. It is specifically trained for agentic clinical reasoning, enabling it to actively retrieve patient-specific evidence from raw EHR tables and consult external medical knowledge. This model excels at long-horizon evidence-seeking behavior in clinical settings, utilizing native tool-call formats for interacting with electronic health records.
Loading preview...
Overview
UCSC-VLAA/ClinSeek-35B-A3B is a 35.1 billion parameter model, fine-tuned from Qwen/Qwen3.5-35B-A3B using supervised learning on ClinSeekAgent trajectories. These trajectories were generated by Claude Opus 4.6, teaching the model to imitate complex evidence-seeking behaviors in clinical reasoning. The model's primary focus is on automating multimodal evidence seeking for agentic clinical reasoning, particularly within Electronic Health Record (EHR) environments.
Key Capabilities
- Agentic Clinical Reasoning: Designed to actively retrieve patient-specific evidence from raw EHR tables and integrate external medical knowledge.
- Tool-Call Integration: Trained to use native tool-call formats (
<tool_call>/<tool_response>) for searching EHRs, treating them as programmable databases. - Enhanced Performance: Achieves a significant improvement over its base model, increasing average F1 score on AgentEHR-Bench from 22.1 to 34.0, a +11.9 point gain.
- Evidence Seeking: Learns a distinct tool-use policy, demonstrated by a substantial increase in free-form SQL calls during evidence retrieval.
Use Cases
- Automated Clinical Evidence Retrieval: Ideal for applications requiring automated search and synthesis of clinical evidence from structured and unstructured data.
- Clinical Decision Support Systems: Can be integrated into systems that assist clinicians by providing relevant patient data and medical knowledge.
- Research in Agentic AI: Serves as a valuable open-source model for researchers exploring agentic behavior and tool use in complex domains like healthcare.
For detailed evaluation scripts and benchmark reconstruction, refer to the ClinSeekAgent repository. The technical report provides further insights into its methodology and performance: ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning.