Chao Huang

I believe AI should learn more by wasting less.

omni-modal LLMsmultimodal reasoningaudio-visual generation

What we see, hear, and say often repeats the same story in different forms. I study how AI can use those overlaps to reason more clearly and spend computation only where it truly needs to.

My work connects multimodal LLMs, audio-visual learning, and diffusion-based generation.

Ph.D. Computer Science, University of Rochester · B.Eng, Nanjing University

Email CV Scholar GitHub

News

Apr 2026 Glad to be recognized among the top 200 reviewers at ICLR 2026, especially meaningful for me in this AI-assisted reviewing era.

Apr 2026 DRIFT accepted to ACL 2026 Findings.

Mar 2026 Successfully defended my PhD dissertation.

Mar 2026 Invited talk on efficient adaptation of foundation models at Adobe Research and Simon Fraser University.

Feb 2026 Three papers accepted to CVPR 2026 (including one at Findings).

Jan 2026 XModBench accepted to ICLR 2026. CAT-V received AAAI 2026 Best Demo Award, Runner-up.

Jan 2026 Joined Tencent Hunyuan as a Research Scientist.

Earlier

Oct 2025 Received the NeurIPS 2025 Scholar Award.

Sep 2025 Three papers accepted to NeurIPS 2025; one to IJCV.

Aug 2025 Selected for ICCV 2025 Doctoral Consortium.

Jun 2025 One paper accepted to ICCV 2025.

Feb 2025 Two papers accepted to CVPR 2025.

Dec 2024 DAVIS received ACCV 2024 Best Paper Award, Honorable Mention.

Jul 2024 Acoustic Primitives accepted to ECCV 2024.

Sep 2023 One paper accepted to NeurIPS 2023.

Feb 2023 One paper accepted to CVPR 2023.

Selected Works

A few projects that define what I care about.

DRIFT

The first method to inject directional reasoning structure into MLLM fine-tuning. Instead of hoping reasoning emerges, we explicitly steer models toward structured, faithful cross-modal inference.

ACL Findings 2026 Paper Project Code

ZeroSep

A zero-shot paradigm shift for audio separation: separate any sound category without ever training on it. Pretrained diffusion priors replace paired training data entirely.

NeurIPS 2025 Paper Project Code

DAVIS

The first generative diffusion approach to visually-guided sound separation, producing high-fidelity audio from diverse real-world mixtures. Best Paper Award at ACCV 2024.

ACCV 2024 → IJCV 2025 Best Paper Award, Runner-up Paper Project Code

20+ papers at NeurIPS, CVPR, ICCV, ECCV, ICLR, IJCV, and more.

View all publications →

Experience

Tencent Hunyuan

Research Scientist

2026 – Present

Omni-modal LLM research and deployment at scale.

AMD Research

Research Scientist Intern

2025

Multimodal reasoning and efficient adaptation of large language models. Led to DRIFT, accepted at ACL Findings 2026.

Meta Reality Labs Research

Research Scientist Intern · Cambridge, UK

2024

Audio-visual learning for cinematic audio highlighting and sound design. Resulted in a CVPR 2025 paper on audio highlighting.

Meta — Codec Avatars Lab

Research Scientist Intern · Pittsburgh

2023

Neural acoustic modeling and spatial audio for human body soundfields. Published at ECCV 2024.

The Chinese University of Hong Kong

Research Assistant

2019 – 2020

3D point cloud processing and non-local denoising methods.

Talks

Mar 2026

Efficient Adaptation of Foundation Models for Multimodal Content Creation Adobe Research · Invited Talk