I build multimodal and multi-agent LLM systems with applications in affective computing, social intelligence, and time series analysis. I will be joining the Melady Lab at USC Viterbi this fall as a Ph.D. student advised by Prof. Yan Liu.
I am a CS (Artificial Intelligence) M.S. graduate from USC Viterbi.
I worked as a graduate researcher at the
Intelligent Human Perception Lab, USC ICT
with Prof. Mohammad Soleymani, focusing on benchmarking and improving multimodal LLMs, and mitigating multimodal hallucinations.
I also worked at the Melady Lab
with Prof. Yan Liu
on multi-agent systems and reinforcement learning.
My first author work, VoxParadox, accepted at ICML 2026! 🔥 🔥
I will be joining the USC Melady Lab at USC Viterbi as a Ph.D. student in Fall 2026!
MoD-DPO accepted at CVPR 2026! 🔥 🔥
Two papers, AVERE and HILA accepted at ICLR 2026! 🔥 🔥
I joined the USC Melady Lab as a graduate researcher, advised by Prof. Yan Liu. Looking forward to explore new directions!
I joined the Intelligent Human Perception Lab at USC ICT as a graduate researcher, advised by Prof. Mohammad Soleymani. Looking forward to the exciting research ahead!
Adversarial benchmark to evaluate language dominance on audio LLMs' paralinguistic understanding, with controlled linguistic-acoustic contradiction (2k clips, 10 tasks). Proposed PCLM, a novel layer fusion strategy improving Audio Flamingo 3 by 48% and Qwen2-Audio by 57% coupled with DPO.
Knowledge corruption attacks against GraphRAG achieving up to 70% attack success rate with GPT-4o-mini. Developed a benchmark with highly inter-connected named-entity graphs and an extensive per-question indexing pipeline.
Fusion-based audio-textual speech emotion captioning with early, late, and X-Norm fusion strategies. X-Norm achieved 20%↑ GPT-evaluated match rate over early/late baselines using prefix-conditioned decoders with a pretrained LLM.
A new method for detecting benchmark leakage by exposing model overfitting to mislabeled or low-confidence examples in SNLI. Identifies dataset outliers via unanimous aggregation of four BERT-based models, human annotation, and clustered Borda ranking with confidence-based safeguards.
More details on CV.