Research

Welcome to the MARGIN Lab research page.
We work on surgical AI, VLMs, explainability, and reasoning-aware perception.

---

🧪 SurgXBench: Explainable Vision-Language Model Benchmark for Surgery

WACV 2026 — Accepted
Jiajun Cheng, Xianwu Zhao, Sainan Liu, Xiaofan Yu, Ravi Prakash, Patrick Codd, Jonathan Katz, Shan Lin

SurgXBench introduces the first explainability-driven benchmark for Vision-Language Models in robotic surgery.
We evaluate general & surgical VLMs for instrument & action recognition, visualize model reasoning using Grad-CAM + causal graphs, and introduce attention-alignment metrics to assess whether models rely on clinically meaningful visual cues.
Results reveal a gap between accuracy and reasoning, motivating the need for more grounded supervision in surgical VLMs.

📄 Paper → https://arxiv.org/abs/2505.10764
💻 Code → https://github.com/jiajun344/SurgXBench-Explainable-Vision-Language-Model-Benchmark-for-Surgery

📌 More Works Coming Soon