|
Decomposing Modality Conflict Mechanism in Vision-Language Models
Status: In progress (pilot study accepted at Actionable Interpretability Workshop at ICML 2025)
TL;DR: This project investigates how foundation models detect and resolve conflicting signals between images and text. We identify decoupled internal mechanisms for conflict detection and resolution.
|
|
Studying Priming Effect in Vision-Language Models
Status: Poster presented at New England Computer Vision (NECV) 2024
TL;DR: This work explores how initial prompts ("primers") can steer VLM behavior even when contradicting visual inputs are present. We investigated whether the priming behavior can be localized to a small set of components within the models.
|
|
Optimal Fusion of Genotype and Drug Embeddings in Predicting Cancer Drug Response
Status: Published in Briefings in Bioinformatics (2024)
TL;DR: We investigate how best to combine gene and drug representations using visible neural networks. Our method improves performance by injecting multiplicative interactions and identifies optimal fusion strategies across biological hierarchies.
|