← Home

Research

I work on efficient multimodal models — vision–language models that have to run under a fixed compute or token budget. A lot of recent work compresses the visual tokens a model sees (merging, pruning, pooling) and reports that average benchmark accuracy holds up. I'm interested in the part that average accuracy hides: whichcapabilities break first under compression, and whether the failure looks like dropped information or confidently wrong information.

Interests

• Visual-token compression for VLMs (merging, pruning, pooling).
• Capability-specific degradation under a fixed token budget.
• Failure modes: omission vs. hallucination, and what causes which.

Code

Selected projects are on the projects page.