Zhenya Liu
Portrait of Zhenya Liu

I am a PhD student in Computer Science at the University of Chicago, advised by Prof. Yuxin Chen.

I am interested in reinforcement learning algorithms that are theoretically grounded and practically useful.

I am currently looking for research internship opportunities across all areas of reinforcement learning. Please feel free to reach out to me at zhenya@uchicago.edu.

Email GitHub

Selected publications

Research projects

2026 · Preprint

Shared Actors Need Not Share Critics: Effects of Value Mismatch in Parallel Reinforcement Learning

Zhenya Liu, Yang Meng, Zhuokai Zhao, Xuefeng Liu, Yuxin Chen

A shared critic can distort sampled learning even when its expected policy gradient remains correct. We identify cross-environment value mismatch as the mechanism and correct it by conditioning only the critic on a logged environment index. Aggregated over all 16 Procgen games, a multihead critic improves normalized held-out return by 40.8% over the shared critic on 600 unseen levels per game.

Project page Code arXiv alphaXiv

Random coverage followed by active refinement

ICML 2026

Active Curriculum Refinement for Reinforcement Learning

Zhenya Liu, Yuxin Chen

PATH turns curriculum learning into active path acquisition on a structured environment DAG. It first explores diverse curriculum paths, then expands successor regions behind newly mastered frontier environments. On 1,000 held-out terrains from a more tractable BipedalWalker subspace, PATH improves solved rate by 97% and mean score by 66% over ACCEL.

Project page OpenReview Code arXiv alphaXiv
Framework from Rethinking Transfer in Continual Learning, organizing transfer around headroom, a persistent carrier, and source selection
Framework overview from the paper (Figure 1).

2026 · arXiv preprint

Rethinking Transfer in Continual Learning: A Replay-Based Realisation

Yang Meng, Zhenya Liu, Zhuokai Zhao, Yuxin Chen

Transfer Selective Replay asks when past experience can genuinely improve a new task, then replays only a compatible source. It operationalizes target headroom, a persistent carrier, and source compatibility through routing and distillation. Across four TRACE-8 and NumGLUE-8 model-scale settings, TSR achieves the best final average accuracy and is the only method with positive backward transfer in every setting.

Code arXiv alphaXiv

2023 · arXiv preprint

The Permutability of σi-Sylowizers of Some σi-Subgroups in Finite Groups

Zhenya Liu, Wenbin Guo

New characterizations of supersoluble finite groups. The paper shows how permutability conditions on σi-sylowizers constrain the structure of finite groups.

arXiv alphaXiv