Robotics Ph.D. candidate at Carnegie Mellon University. I make frozen vision-language models do new things without touching their weights — treating adaptation as an inference-time problem, across object detection and medical imaging.
🎯 On the research-scientist job market (graduating late 2026) · 📄 Résumé · 🌐 ggare-cmu.github.io
- Adapt — prompt optimization for few-shot detection, image-conditioned attribute selection, and activation-space reward models — state-of-the-art with black-box VLMs.
- Understand — when and why multimodal models generalize (class-taxonomy distance, natural experiments, the question-first paradox).
- Deploy — interpretable lung-ultrasound biomarkers and clinical ML: 5 ISBI papers, a Medical Image Analysis journal paper, and 4 U.S. patents.
- DetPO — In-Context Learning with Multi-Modal LLMs for Few-Shot Object Detection · ECCV 2026 — code · page · arXiv
- Attributes Should Come from Images, Not Class Names · ECCV 2026 Workshop (Oral) — page · arXiv
- Activation Reward Models for Few-Shot Model Alignment · ACL Findings 2026 — paper
- LCA-on-the-Line — Benchmarking OOD Generalization with Class Taxonomies · ICML 2024 (Oral) — paper