An Exam for Active Observers
We introduce ActiveVision, a benchmark of 17 tasks that require repeated visual observation, revealing a substantial gap between frontier multimodal models and human performance.
Jerry (Jiarui) Zhang
Ph.D. Student in Computer Science · University of Southern California
I am a Computer Science Ph.D. student at USC, advised by Willie Neiswanger. My research focuses on multimodal learning and reasoning. Previously, I received my bachelor's degree in electrical engineering from Tsinghua University.
I’m looking for industry jobs starting in Spring/Summer 2027. Please reach out to me by email!
Research
Multimodal learning and reasoning.
MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs
We show that MLLMs often know where to look but fail to perceive small details, and that attention-guided cropping improves performance without training.
Euclid: Supercharging Multimodal LLMs with Synthetic High-Fidelity Visual Descriptions
Euclid studies geometric low-level visual perception in multimodal LLMs and introduces Geoperception, design-space analyses, and a model family with stronger LLVP abilities.
Exploring Perceptual Limitations of Multimodal LLMs on Small Visual Objects
We reveal that object quality, size, distractors, and location independently affect MLLMs' ability to perceive small visual objects.
MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning
A comprehensive benchmark, MARVEL, that evaluates multimodal large language models' abstract reasoning abilities and reveals significant performance gaps between humans and state-of-the-art MLLMs.