# Projects: Peter Li

- In the room: <https://peterli.space/#projects>
- Index: <https://peterli.space/llms.txt>

## BODHI
- Role: Researcher
- Year: 2026
- Overview: LoRA distillation of clinical epistemic behavior into MedGemma-27B, under review at NeurIPS 2026 (Massachusetts Institute of Technology)
- Image: https://peterli.space/bodhi-thumb.png
- Image shows: Mean rubric score across difficulty tiers comparing base model, BODHI wrapper, and LoRA-adapted variants.

### Context

BODHI started as a prompt-engineering wrapper for clinical language models. The goal was to move beyond runtime prompting toward genuine model adaptation, building something that could internalize clinical reasoning rather than mimic it at inference time.

### Process

Built a four-step pipeline that distills clinical epistemic behavior (uncertainty acknowledgment, active inquiry, scope bounding, red-flag recognition) into model weights via LoRA. A structured chain-of-thought protocol generates demonstrations, expert rubrics filter quality, and the adapter internalizes the behavior. Applied to MedGemma-27B on ~3,100 filtered traces, evaluated on HealthBench.

### What Changed

First-author paper, "Ask Before You Answer: Distilling Clinical Epistemic Calibration into Model Weights," under review at NeurIPS 2026. Active clinical inquiry rose from 17.5% to 45.6% without degrading clinical quality. Code, adapters, and a seven-dimension behavioral evaluation framework are open-source.
