I am a third-year PhD student at the University of Southern California (USC). I work primarily with Prof. Konstantinos Psounis (he is fantastic!). My recent research focuses on LLM/VLM Post-Training (Reinforcement Learning, Distillation) and Evaluation, Reasoning & Agentic AI. I’m always open to discussion, collaboration, and mentorship!
Before my PhD, I completed my Bachelor's with double majors in CS and Statistics at Ohio State University from 2019-2022, working with great Prof. Wei-Lun Chao and Dr. Hong-You Chen. I have also spent some wonderful time at CCVL lab at Johns Hopkins University, supervised by Bloomberg Distinguished Prof. Alan Yuille.
I'm currently looking for Spring 2027 and Summer 2027 research internship opportunities, please feel free to reach out at jikezhon@usc.edu if you see a fit.
Full list of publications can be found on my Google Scholar page.
We identify why student agents surpass their teachers during distillation, then lose those gains. Our success-preserving method, SPOD, improves final success by up to 15.2 percentage points over standard OPD and 20 points over the teacher.
We expose shortcuts that inflate apparent theory-of-mind reasoning performance. On shortcut-resistant datasets, reinforcement fine-tuning with explicit reasoning improves generalization, higher-order reasoning, and robustness to counterfactuals.
We benchmark how multimodal agents track and reconcile evolving, information-dense visual artifacts across professional workflows. DSV-Mem exposes state reconstruction as a key memory bottleneck and shows the value of explicit state tracking.
We evaluate agents that use tools and code to manipulate visual evidence across 13 tasks and 1,215 examples. Testing 22 multimodal models reveals substantial gaps in agentic visual reasoning; TIR-Bench also informs Qwen3.5 and Qwen3.6 evaluations.
We show that explicit thinking is not always beneficial during visual reinforcement fine-tuning: direct-answer training can match or outperform it on perception tasks while reducing computation. We further explore how models learn whether to think based on their capabilities and task complexity.
Applied Scientist Intern · AGI Foundations Group
Foundation-model post-training and stateful memory for long-horizon multimodal agents.
Research Intern · Cognition Group
Robust LLM/MLLM post-training for theory-of-mind reasoning.
Member of Technical Staff · Infra Data Science Group
Software Engineer Intern
Software Engineer Intern
I’m a big aviation fan. Check out my photographs below or on JetPhotos and Airliners.
I also enjoy classical piano and opera, especially those by Franz Liszt (Hungarian Rhapsody) and Bellini (I Puritani), though the one I play most often currently is Ballade No. 1.
