Xiangchi Yuan

Hi! I am a Ph.D. student in Computer Science at Georgia Institute of Technology, advised by Prof. Wenke Lee.

I work on training large foundation models to reason and solve complex tasks, with a focus on agentic systems, reinforcement learning, and long-context modeling. I am fortunate to collaborate with wonderful researchers across academia and industry.

Previously, I received my M.S. in Computer Science from Brandeis University and my B.Eng. from UESTC and the University of Glasgow.

Email/CV/Google Scholar/GitHub/LinkedIn

Research

My research interests include agentic reasoning, SFT/RL post-training, and long-context foundation models. I am especially interested in making reasoning systems more capable, efficient, and reliable.

News

2026 New preprint: Priority-Preserving Advantage Aggregation for Multi-Reward Policy Optimization.
2026 MIFO and SafeCoDe at EMNLP Main; MetaState at COLM; Dynamic Top-p MoE at ICML.
2026 RAM and Curriculum Evaluation at ACL Main, MTMCS-Bench at ACL Findings, and SwiReasoning at ICLR.
May 2026– Research intern at ByteDance Seed, San Jose, working with Dr. Ming Tu and Dr. Lu Lu on multimodal RL and agentic audio.
2025 Papers at EMNLP, ACL (including an Oral), ICML, and ICLR.
May–Aug 2025 Research intern at Adobe Research, San Jose, hosted by Dr. Xiang Chen. This reasoning post-training work is published at EMNLP 2026.

Experience

ByteDance Seed · San Jose, CA · Research Intern · May 2026–Present

Core contributor to the multimodal RFT/RL recipe for the audio understanding component of the Seed (Doubao) flagship model, including data-mixture design, advantage aggregation, and process-reward optimization with PPO/VAPO. Improved audio-text understanding by 5.1% on average across internal and public benchmarks (MMAU, MMSU), outperforming previous SOTA Gemini 3.1 Pro in this domain.

Built an agentic audio framework (thinking-with-audio) for training and evaluation from scratch for long-form speaker diarization, reducing DER from 55.7% to 11.3% on the Seed Omni model.

Adobe Research · San Jose, CA · Research Intern · May–August 2025

Developed an efficient post-training framework for reasoning agents that dynamically interleaves RL and SFT while mitigating cross-stage forgetting, achieving state-of-the-art reasoning performance with only 7.8% of the training data.

Published at EMNLP 2026 and filed a U.S. patent.

Preprints

Selected Publications

Full list →

* denotes equal contribution. My name is highlighted.

Loading publications…

Selected Collaborations

Full list →

Contact

If you would like to chat about research, career, or life, feel free to reach out.