work

Working on search, reasoning, and verification for Large Language Models (LLMs). Recent work focuses on test-time verifiers that monitor and intervene in agent trajectories at runtime, as well as the Deep Research (DR) paradigm, where LLM systems perform structured search over large corpora.

Benchmarked vision-language models on inferring the message in rhetorical and atypical imagery used in advertisements, through three novel tasks. Found that VLMs struggle with reasoning about atypicality, but simple methods enable atypicality-aware verbalization and improve rhetorical image understanding.

Worked on modelling and simulating human behavior at scale from Youtube Videos. Trained multimodal LLM architectures that integrated video and textual representations, achieving significant performance gains and surpassing GPT-4 on multiple behavior prediction tasks. Published at ICLR 2024, with a spotlight presentation.