Adversarial Paralinguistic Benchmark
28k+ examples with mismatched text-audio cues to diagnose transcript over-reliance in speech-LLMs (work in progress).
Hi, I'm Jiacheng Pang.
I build multimodal and multi-agent LLM systems with applications in affective computing and social intelligence.
I am currently a CS(AI) Master's student at USC Viterbi (graduating in December 2025).
I am also working as a research assistant at the
Intelligent Human Perception Lab, USC ICT
with Prof. Mohammad soleymani and
Ashutosh Chaubey, where I focus on building and evaluating multimodal LLMs for paralinguistic understanding.
I am also working at the Melady Lab at USC Viterbi
with Prof. Yan Liu and Wei Yang
on multi-agent collaboration and reinforcement learning.
Areas of Interests:
Multimodality
Multi-agent
LLM Evaluation & Alignment
Multimodality
Multi-agent
LLM Evaluation & Alignment
Education
- University of Southern California
M.S. Computer Science (Artificial Intelligence), GPA: 3.91/4.00, (Jan. 2024 - Dec. 2025) - Northeastern University
B.S. Computer Science, GPA: 3.74/4.00, (Sep. 2018 - May 2022)
Honors: Dean's List of all semesters, magna cum laude
Selected Projects
Fusion Based Speech Emotion Captioning
Encoder-decoder captioning model with audio-textual fusion, enhancing match rate by 20% with X-Norm vs early/late fusion.
Benchmark Leakage Detection
Detection through exposing model overfitting to mislabeled or low-confidence examples, identified with majority voting among model judges.
More details on CV...
Publications
-
VoxParadox: Speech LLMs Hear the Words, Not the VoiceSubmitting to ICML 2026
-
A Survey of Reasoning and Agentic Systems in Time Series with Large Language Models.
More under double-blind review at ICLR.
Experience
Graduate ResearcherJune 2025 - Present · Los Angeles, USAGraduate ResearcherJune 2025 - Present · Los Angeles, USASoftware EngineerSept. 2022 - Nov. 2023 · Yinchuan, ChinaData Science Researcher InternJan. 2021 - July 2021 · Beijing, China