Heng Dong
About Me
I am a Ph.D. Student advised by Prof. Chongjie Zhang and Prof. Yi Wu at the Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, led by Prof. Andrew Yao.
The goal of my research is to endow agents with superhuman intelligence, which I believe can be achieved through learning from interactions using modern models. My biography is here.
Research Interests
- Learning from Interactions: RL, Robot Control & Design, Multi-Agent
- Modern Models: LLMs & VLMs, Diffusion Models, Flow Models
Education
- [2020 - 2025] Ph.D. in Artificial Intelligence, Tsinghua University
- [2016 - 2020] B.S. in Computer Science, University of Science and Technology of China
News
- [May 2025] Our paper about enhancing the decision-making ability of LLMs is accepted to ICML 2025.
Publications
-
ICML 2025TL;DR We propose a novel LLM-based Actor-Critic framework that enhances LLMs' decision-making through long-term action evaluations and efficient policy improvements.
-
arXiv 2025
TL;DR We train LLMs to generate better agent explanations using RL with feedback from generative models, outperforming standard RLHF methods. -
AAMAS 2025TL;DR We show that diffusion models can reconstruct global states in decentralized partially observable multiagent systems, with approximation errors leading to deviations that can be bounded for convergence to the true state.
-
ICLR 2024TL;DR We propose to design multi-cellular robots in a coarse-to-fine manner and leverage hyperbolic embeddings for implementation.
-
ICML 2023TL;DR We exploit the structure of the design space in robot design problems with symmetry characteristics and generate robots with high performance more efficiently.
-
NeurIPS 2022TL;DR A muscle synergy inspired RL framework that groups actuators into synchronised modules for efficient control of high-DoF robots, demonstrating superior efficiency and generalizability on complex morphologies like Humanoids++ and UNIMALs.
-
arXiv 2021TL;DR We identify that fragile cooperation in sequential social dilemmas stems from second-order conflicts in incentive mechanisms, and propose a homophily-based learning framework that achieves stable cooperation in public goods and tragedy of the commons scenarios.
-
ICLR 2021
TL;DR A decomposed policy gradient method that outperforms state-of-the-art multi-agent RL algorithms by enabling efficient off-policy learning and better credit assignment. -
ICML 2020
TL;DR A role-based MARL framework where agents autonomously learn specialized roles, improving performance on StarCraft II micromanagement.
Most of my research is about robotics and agents. Some papers are highlighted.
Services
Conference Reviewer
- [2022 - present]
Annual Conference on Neural Information Processing Systems (NeurIPS) - [2022 - present]
International Conference on Machine Learning (ICML) - [2022 - present]
International Conference on Learning Representations (ICLR) - [2025 - present]
Association for the Advancement of Artificial Intelligence (AAAI)
Teaching Assistant
- [Fall, 2021]
Artificial Intelligence: Principles and Techniques, IIIS, Tsinghua - [Spring, 2022]
Reinforcement Learning, IIIS, Tsinghua
Powered by Jekyll and Minimal Light theme. Also some ideas borrowed from Jon Barron and Junyu Zhang