Chenghao Liu

Master's Student, Peking University

Email: chliu [at] stu [dot] pku [dot] edu [dot] cn

Last updated: October 20, 2026

Biography

I am currently pursuing my master's degree at Peking University, advised by Prof. Songfang Huang. I received my bachelor's degree from Huazhong University of Science and Technology.

My research interests mainly focus on: 1) Language Model Architecture and Training/Inference Methods: designing model architectures and training/inference techniques for discrete-token language models and continuous diffusion language models to improve task-specific performance or general capabilities; 2) Agentic Systems: memory, workflow, and data flywheel for LLM-based agents.

If you share similar research interests, feel free to reach out.

I expect to graduate in June 2027 and am currently seeking full-time opportunities. Please feel free to reach out.

News

[Oct. 2026] Our paper MORPH has been accepted to TMLR! MORPH introduces a training pipeline spanning pre-training, CPT, SFT, and RL. It adapts a pretrained discrete-token language model into a continuous diffusion language model for large-scale vector retrieval via flow matching. [Paper]
[Oct. 2026] We released CASE, a framework for cost-aware stopping in long-video agents. [Paper] · [Code]
[Oct. 2026] Serving as a reviewer for ICLR 2027 and ICASSP 2027.
[Sep. 2026] Our paper PACE, on phase-aware action chunk execution for robot policies, has been accepted to NeurIPS 2026! [Paper]
[Sep. 2026] X-Rec Technical Report is now available! X-Rec is a DiT-based generative recommendation model that uses flow matching in continuous item embedding space and has been deployed on TikTok. [Report]
[Aug. 2026] We released MemOPD, aligning memory states for on-policy distillation of long-horizon agents. [Paper] · [Code]
[Aug. 2026] Serving as a senior program committee member for AAAI 2027.
[Jun. 2026] Joined Baidu’s BMU pre-training team (i.e., ERNIE pre-training team) as a research intern, working on model architectures for pre-training and training dynamics.
[Jan. 2026] Our paper MSNav, on using an LLM agent for zero-shot vision-and-language navigation, has been accepted to ICASSP 2026! [Paper] · [Code]
[Dec. 2025] Joined TikTok’s Content Intelligence Team & Feed Quality Team as a research intern, working on VLM post-training and continuous diffusion language models for generative retrieval and recommendation.
[Nov. 2025] Our paper TTF-VLA, on improving vision-language-action model inference through training-free temporal token fusion, has been accepted to AAAI 2026! [Paper] · [Code]
[Aug. 2025] Serving as a program committee member for AAAI 2026.
[Feb. 2025] Joined Samsung R&D China’s Advanced Research Lab as a research intern, working on VLM and VLA architectures.
[Sep. 2024] Started my master’s studies at Peking University and joined XLab as a student researcher, advised by Prof. Songfang Huang.
[Jun. 2024] Graduated from Huazhong University of Science and Technology with the Outstanding Graduate honor.

Experience

	Baidu - Basic Model R&D Dept. - Pre-training Algorithm Team Research Intern, Jun. 2026 - Present Focus: ERNIE Series Models, LLM Pre-training Architecture, Training Dynamics
	ByteDance - TikTok - Content Intelligence Team & Feed Quality Team Research Intern, Dec. 2025 - Jun. 2026 Focus: VLM Post-training, Continuous Diffusion Language Models
	Samsung R&D China - Advanced Research Lab Research Intern, Feb. 2025 - Apr. 2025 Focus: Model (VLM, VLA) Architecture
	Peking University - XLab Student Researcher, Sep. 2024 - Present Focus: Model (LLM, VLM, VLA) Architecture, Token Efficiency, Agents

Education

Peking University

School of Advanced Manufacturing and Robotics

M.Eng., Sep. 2024 - Jun. 2027 (expected)

Huazhong University of Science and Technology

School of Cyber Science and Engineering

B.Eng., Sep. 2020 - Jun. 2024

Selected Publications [Full List]

X-Rec Technical Report

Chenghao Liu (Contributor, TikTok Content Intelligence Team)

Tech Report [paper]

PRISM: Probing and Repairing Internal States in Language Model Pretraining

Chenghao Liu, Junyuan Shang, Tian-Chi Ma, Zhiyuan Liu, Shuohuan Wang, Dianhai Yu, Songfang Huang

Preprint

CASE: Cost-Aware Stopping for Efficient Long-Video Agents

Yiming Du, Chenghao Liu, Zhiyuan Liu, Fangxing Zheng, Zhao Wang, Junnan Nie, Songfang Huang

Preprint [paper] [code]

MORPH: Generative Retrieval via DiT with Metric-Ordered Sequence Training and HPPO

Chenghao Liu, Yu Zhang, Zhongtao Jiang, Kun Xu, Zhenwei An, Renzhi Wang, Zhao Wang, Jiachen Zhang, Yuxiao Zhang, Kun Xu, Songfang Huang

TMLR 2026 [paper]

PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking

Junnan Nie, Jiayi Li, Chenghao Liu, Junyi Lao, Jiachen Zhang, Tianle Zhang, Liang Lin, Songfang Huang

NeurIPS 2026 [paper]

TTF-VLA: Temporal Token Fusion via Pixel-Attention Integration for Vision-Language-Action Models

Chenghao Liu, Jiachen Zhang, Chengxuan Li, Zhimu Zhou, Shixin Wu, Songfang Huang, Huiling Duan

AAAI 2026 [paper] [code]

MSNav: Zero-Shot Vision-and-Language Navigation with Dynamic Memory and LLM Spatial Reasoning

Chenghao Liu, Zhimu Zhou, Jiachen Zhang, Minghao Zhang, Songfang Huang, Huiling Duan

ICASSP 2026 [paper] [code]

Beyond Research

Non-Negative-Sum Games. I am interested in non-negative-sum environments (i.e., zero-sum and positive-sum games), from financial markets to strategy games. In my spare time, I trade and also enjoy games such as Texas Hold'em and Mahjong.

My interests cover almost all tradable assets, including U.S. stocks, Hong Kong stocks, Chinese A-shares, cryptocurrencies, spot markets, futures, derivatives, options, prediction markets, and more. I am a firm believer in Kondratiev waves and agree with the value-investing philosophy of buying when assets are undervalued and selling when valuations are too high. I also recognize the value of data-driven high-frequency technical trading, such as various liquidity strategies and spread arbitrage strategies.

I enjoy analyzing these non-negative-sum environments and trying to find an edge through strategy and reasoning to earn real returns. Recently, I have also been exploring LLM applications in trading, including factor mining, automated trading systems, news-driven strategies, and the analysis of public opinion and market sentiment, among others.

Current Aspiration. My medium- to long-term goal is to accumulate capital through earned income and trading profits, then gradually achieve FIRE (Financial Independence, Retire Early) through compounding. Ultimately, I hope to “buy back my time and freedom to do what I truly want to do”.
Video Games. I also enjoy video games, including single-player games on platforms such as PlayStation, Nintendo Switch, and Steam, as well as cooperative multiplayer games. My favorite game is The Legend of Zelda: Breath of the Wild—the hundreds of hours I spent playing it are among the happiest moments of my life. For multiplayer games, I particularly enjoy League of Legends, which is probably the game I've spent the most time playing.

If you happen to share similar interests and aspirations, feel free to reach out.