Wu, J. (CSE) – From Reliable Reasoning to Self-Improving Agents: Toward General Medical Intelligence

Large language models (LLMs) are increasingly applied to medicine, and much of their recent progress comes from training them to reason in longer chains of thought. Reasoning longer, however, neither makes each step correct nor lets a model obtain information it has not been given. General medical intelligence therefore requires two further abilities. The first is reliable reasoning, in which every step is consistent with medical knowledge and with what the inputs actually show. The second is agentic ability, in which the model obtains the information a decision requires through multi-step interaction with patient records and tools, and improves with experience.

This proposal progresses from the first ability to the second, addressing problems that are general to language models in a setting where errors carry a high cost. MedReason makes medical reasoning factually reliable by generating chain-of-thought data grounded in a medical knowledge graph; fine-tuning on the resulting 32,682 examples improves both instruction-tuned and reasoning LLMs, by up to 7.7 percentage points. From Seeing to Thinking finds that 86.9% of the errors of Qwen3-VL-8B on visual math stem from perception rather than reasoning, and decouples post-training into perception, textual reasoning, and visual reasoning stages; this curriculum over capabilities outperforms merged training and complements curricula over sample difficulty.

Building on these foundations, our ongoing work, ClinSeekAgent, turns the model into an agent that seeks evidence from raw longitudinal health records and medical imaging tools over multiple rounds. We then propose step-level self-improvement, in which the agent diagnoses its own trajectories and trains on the steps it has not yet mastered instead of imitating whole teacher trajectories. The progression moves from reasoning over given information to agents that acquire evidence and learn from their own experience, as steps toward general medical intelligence.

Event Host: Juncheng Wu, PhD Student, Computer Science & Engineering 

Advisor: Yuyin Zhou 

Zoom: https://ucsc.zoom.us/j/99581490205?pwd=MEtt3rGG2q0tfW1vLDbKd8ftmu7G7Y.1

Passcode: 124896