Henglin Liu

Henglin Liu (刘恒霖)

Hi, I'm Henglin Liu. I am currently a master’s student at Tsinghua University, under the supervision of Prof. Xiu Li and Prof. Xiangyang Ji. I obtained my B.Eng in Artificial Intelligence at Xiamen University in 2024.

My current research interests includes:

Multimodal large language model/Reward Model

A focus on enhancing the core capabilities of large multimodal models in tasks such as quantitative computation, logical reasoning, and structured analysis.
A key application is training reward models to predict human-aligned evaluation scores, which supports training and optimizing non-verifiable reinforcement learning tasks.

Reinforcement Learning on Large Models

Aligning Diffusion models, Large Language Models with specific human preferences and demands, with techniques like Reinforcement Learning from Reward (RLHF/RLAIF/RLVR).
Exploring robust fine-tuning "recipes" within the RL framework to ensure that pre-trained capabilities are preserved while desired, human-aligned skills are effectively amplified.

Email / Scholar / Github

Publications:

Reward Modeling with multimodal vision-language model

Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment
Henglin Liu, Nisha Huang, Chang Liu, Jiangpeng Yan, Huijuan Huang, Jixuan Ying, Tong-Yee Lee, Pengfei Wan, Xiangyang Ji
The Association for the Advancement of Artificial Intelligence (AAAI), 2026
Poster

A framework enhances the alignment between models and human aesthetic judgment through auxiliary hierarchical description generation tasks.

Robust Reinforcement Learning with Diffusion Model

DiverseGRPO: Mitigating Mode Collapse in Image Generation via Diversity-Aware GRPO
Henglin Liu, Huijuan Huang, Jing Wang, Chang Liu, Xiu Li, Xiangyang Ji
Under Review
Project

DiverseGRPO addresses mode collapse in image generation by introducing a diversity-aware reward mechanism, enhancing the variety of generated images while maintaining quality.

GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping
Jing Wang, Jiajun Liang, Jie Liu, Henglin Liu, Gongye Liu, Jun Zheng, Wanyuan Pang, Ao Ma, Zhenyu Xie, Xintao Wang, Meng Wang, Pengfei Wan, Xiaodan Liang
Under Review
Project

GRPO-Guard mitigates implicit over-optimization in FlowGRPO, preventing degradation of image quality. Watch our video presentation for a quick summary of our key contributions.

Controllable Image Generation

MaTe: Images Are All You Need for Material Transfer via Diffusion Transformer
Nisha Huang, Henglin Liu, Yizhou Lin, Kaer Huang, Chubin Chen, Jie Guo, Tong-yee Lee, Xiu Li
International Conference on Computer Vision (ICCV), 2025
Paper

A streamlined diffusion framework that eliminates textual guidance and reference networks for material transfer.

ICE: Intercede Concept Erasure in Text-to-Image Diffusion Models
Yizhou Lin, Nisha Huang, Kaer Huang, Henglin Liu, Yiqiang Yan, Jie Guo, Tong-Yee Lee, Xiu Li
ACM Multimedia (MM), 2025
Paper

ICE introduces a novel approach to selectively erase specific concepts from text-to-image diffusion models, enhancing user control over generated content.

Vertical Domain Foundation Model

Research on Temporal Mask Modeling Algorithm for Seismic Wave
Diting Group
Bachelor's Thesis
News Report

Pre-training and fine-tuning of foundation models for seismic wave data, by temporal masked autoencoders (MAE), enhancing the accuracy of seismic data interpretation and analysis.

Advancing financial engineering with foundation models: progress, applications, and challenges
Liyuan Chen, Shuoling Liu, Jiangpeng Yan, Xiaoyu Wang, Henglin Liu, Chuang Li, Kecheng Jiao, Jixuan Ying, Yang Veronica Liu, Qiang Yang, Xiu Li
Engineering
Paper

A comprehensive survey on the applications of foundation models in financial engineering, discussing current progress, practical applications, and future challenges.