About

I am Hao Liang (梁昊), a Ph.D. candidate at Center for Data Science, Peking University, jointly affiliated with Zhongguancun Academy. I am fortunate to be supervised by Prof. Wentao Zhang and Prof. Bin Dong, and to work closely with Prof. Bin Cui and Prof. Weinan E.

Prior to this, I received my bachelor’s degree from Beijing Institute of Technology, where I was awarded the Xu Teli Scholarship (the highest honor of BIT) and the National Scholarship. I also visited the University of Oxford, working with Prof. Ismail Ilkan Ceylan and Prof. Michael Bronstein.

News

[May. 2026] Honored to receive the 2026 President’s Scholarship at Peking University for the second consecutive year.
[Apr. 2026] Our DataFlex Technical Report ranked #1 on the Hugging Face Daily Papers leaderboard for that day.
[Dec. 2025] Our DataFlow Technical Report ranked #1 on the Hugging Face Daily and Weekly Paper Charts!
[Sep. 2025] We won 1st place in the ICML 2025 SeePhys Challenge!
[Jun. 2025] Honored to receive the President’s Scholarship at Peking University.

My research focuses on Data-Centric AI along four directions. For broader context, see our surveys on text-centric and multimodal perspectives.

Data Attribution & Data–Model Interaction (primary focus) — Understanding how individual data shapes model behavior through attribution, and closing the loop between data and models during training via dynamic selection, mixture, and reweighting. This is the core idea behind DataFlex, our data-centric training framework; a conceptual overview appears in Towards Next-Generation LLM Training: From the Data-Centric Perspective.
Data Agents & Principled Data Synthesis — Autonomous agents that curate, transform, and manage data intelligently, together with data synthesis grounded in insight and theory rather than heuristics. Examples include DataFlow-Skills and Text2SQL-Flow.
Data Infrastructure — Scalable systems and pipelines for data preparation and data–model iterative training at scale. The open-source stack is centered on DataFlow.
Data for Science — Curating scientific corpora for training and evaluation, including mathematics and formal-verification data (e.g., Lean). Representative work includes MathScape, MM-Verify, and Let’s Verify Math Questions Step by Step.

I have published 9 first-author / co-first-author papers at CCF-A venues and received the Sa Shixuan Best Student Paper Award at NDBC.

DataFlow — Lead designer of this open-source LLM data-preparation framework (“Generate, Clean, and Prepare LLM Data, All-in-One”). It provides operator-based pipelines for data synthesis, cleaning, and evaluation across domains, and won 1st place in the ICML SeePhy Challenge and 1st place in the Zhiyuan LIC Challenge.
DataFlex — Lead designer of this open-source data-centric training framework built on LLaMA-Factory (“Data Select · Mix · Reweight — Right in the LLM Training Loop”), enabling dynamic data selection, mixture, and reweighting inside the training loop.
LLaMA-Factory — Contributed to the data module design.
CAMEL — Integrated DataFlow into CAMEL’s data pipeline.

President’s Scholarship, Peking University
Industrial Bank Scholarship (兴业奖学金), Peking University
Xu Teli Scholarship (徐特立奖学金), Beijing Institute of Technology (Highest Honor)
National Scholarship, Beijing Institute of Technology
Sa Shixuan Best Student Paper Award (萨师煊优秀学生论文奖), NDBC