top of page

About Me

Dr. Yaodong Yang is the deputy director of Centre for AI Safety and Governance at the Institute for AI, Peking University. Before joining Peking University, he was an assistant professor at King's College London. He studies AI alignment, reinforcement learning, game theory and multi-agent systems, aiming to develop methods that can help achieve intelligent decision making, strategic interaction and human value alignment for the coming AGI era. He has maintained a track record of more than 100 publications at top conferences (NeurIPS, ICML, ICLR) and top journals (Nature Machine Intelligence, AIJ, JMLR, T-PAMI, National Science Review, etc), along with the ICCV'23 best paper award initial list, the CoRL'20 best system paper award, and the AAMAS'21 best blue-sky paper award. He has also been awarded ACM SIGAI China Rising Star and World AI Conference (WAIC'22) Rising Star. He holds a Ph.D. degree from University College London (nominated by UCL for ACM SIGAI Doctoral Dissertation Award), an M.Sc. degree from Imperial College London and a Bachelor degree from University of Science and Technology of China.

杨耀东博士,北京大学人工智能研究院研究员(博导)、AI安全与治理中心执行主任。国家高层次留学人才计划、国家高层次青年人才项目、中国科协青年托举计划、北大博雅青年学者获得者。重点研究具身多智能体系统构建、博弈交互与价值对齐等问题,科研领域包括强化学习、博弈论和多智能体系统。本科毕业于中国科学技术大学,随后在伦敦帝国理工大学、伦敦大学学院获得硕士及博士学位(论文获学校唯一提名ACM SIGAI 优博奖)。回国前曾于伦敦国王大学信息学院任助理教授。发表AI领域顶会顶刊论文一百余篇,谷歌引用五千余次,主持国自然、科技部、市科委、校企联合实验室等科研课题二十余项。曾获国际计算机视觉会议ICCV’23最佳论文奖入围、机器人学习会议CoRL’20最佳系统论文奖、多智能体系统会议AAMAS’21最具前瞻性论文奖、世界人工智能大会WAIC’22云帆奖璀璨明星、ACM SIGAI China新星奖。相关工作曾被央视一套《焦点访谈》央视四套《深度国际》栏目Financial TimesMIT Tech Review报道






   强化学习开源项目(Show me the code, not the story~


Recent News


Valse 2024年度进展报告:从偏好对齐到价值对齐与超对齐



Three papers get accepted at ICML 2024

  1. SINSIGHT: End-to-End Neuro-Symbolic Visual Reinforcement Learning with Language Explanations

  2. Safe Reinforcement Learning using Finite-Horizon Gradient-based Estimation

  3. Planning with Theory of Mind for Few-Shot Adaptation in Mixed-motive Environments


We, alogn with Yoshua Bengio, Stuart Russell, Geff Hinton and Chinese decision makers signed Beijing Declaration on AI Safety.



Five papers get accepted at ICLR 2024 & one paper on TPAMI.

  1. Spotlight (5%) CivRealm: A Learning and Reasoning Odyssey for Decision-Making Agents

  2. Spotlight (5%) Maximum Entropy Heterogeneous-Agent Reinforcement Learning

  3. Spotlight (5%) Safe RLHF: Safe Reinforcement Learning from Human Feedback

  4. SafeDreamer: Safe Reinforcement Learning with World Models

  5. Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game

  6. PAMI  ASP: Learn a Universal Neural Solver


Three papers get accepted at AAAI 2024.

  1. STAS: Spatial-Temporal Return Decomposition for Multi-agent Reinforcement Learning

  2. Oral (7%) ProAgent: Building Proactive Cooperative AI with Large Language Models

  3. A Perspective of Q-value Estimation on Offline-to-Online Reinforcement Learning


Two top journals get accepted!

  1. PAMI Bi-DexHands: Towards Human-Level Bimanual Dexterous Manipulation

  2. JMLR Heterogeneous-Agent Reinforcement Learning



We release AI Alignment Survey and Alignment Resource Website. 


Our paper won the best paper initial list (17/8260) at ICCV 2023!


Six papers get accepted at NeurIPS 2023.


  1. Multi-Agent First Order Constrained Optimization in Policy Space

  2. Hierarchical Multi-Agent Skill Discovery

  3. Policy Space Diversity for Non-Transitive Games

  4. Team-PSRO for Learning Approximate TMECor in Large Team Games via Cooperative Reinforcement Learning

  5. BeaverTails: A Human-Preference Dataset for LLM Harmlessness Alignment

  6. Safety Gymnasium: A Unified Safe Reinforcement Learning Benchmark


Two papers get accepted at JMLR and TMLR.

  1. JMLR MARLlib: A Scalable and Efficient Multi-agent Reinforcement Learning Library 

  2. TMLR  JiangJun: Mastering Xiangqi by Tackling Non-Transitivity in Two-Player Zero-Sum Games


TorchOpt is now officially part of PyTorch Ecosystem!






Four papers get accepted at ICML 2023.

  1. Regret-Minimizing Double Oracle for Extensive-Form Games

  2. MANSA: Learning Fast and Slow in Multi-Agent Systems

  3. A Game-Theoretic Framework for Managing Risk in Multi-Agent Systems

  4. GEAR: A GPU-Centric Experience Replay System for Large Reinforcement Learning Models


Invited talk given:


Slides: Aligning safe decision in open-ended world.



One paper gets accepted at Artificial Intelligence Journal

Safe Multi-Agent Reinforcement Learning for Multi-Robot Control

We propose the first safe cooperative MARL method.


Two ICRA papers, One ICLR paper got accepted.

ICRA'23: End-to-End Affordance Learning for Robotic Manipulation

We take advantage of visual affordance by using the contact information generated during the RL training process to predict contact maps of interest.

ICRA'23: GenDexGrasp: Generalizable Dexterous Grasping

A versatile dexterous grasping method that can generalize to unseen hands.


A new policy diversity measure is proposed that suits game AI settings.


One paper gets accepted at Autonomous Agents and Multi-Agent Systems (Springer)

Online Markov Decision Processes with Non-oblivious Strategic Adversary

We study the setting of online MDP where the adversary is smart where it can change its policy accordingly to the learning agent's behavior.


One paper gets accepted at AAMAS 2023

Is Nash Equilibrium Approximator Learnable ?

We prove that Nash Equilibrium is agnostic-PAC learnable. 


We have won the 1st place at NeurIPS 2022 MyoChallenge!

This competition is about learning contact-rich manipulation using a musculoskeletal hand, e.g., Die Rotation.


Our paper gets accepted at National Science Review [IF-23]

On the complexity of computing markov perfect equilibrium in general-sum stochastic games

We prove the complexity of computing Nash Equilibrium in Markov games are PPAD-Complete.


Three multi-agent RL papers get accepted at AAAI 2023.


Mutli-agent RL:




Talk is given at Airs in Air.


Game Theoretical Multi-Agent Reinforcement Learning.


Talk is given at 2022.

A General Solution Framework to Cooperative MARL.


Seven papers got accepted at NeurIPS 2022.

Preference-based RL:

Meta-Reward-Net: Implicitly Differentiable Reward Learning for Preference-based Reinforcement Learning


A Theoretical Understanding of Gradient Bias in Meta-Reinforcement Learning

Safe RL:

Constrained Update Projection Approach to Safe Policy Optimization

Cooperative Games:

Multi-Agent Reinforcement Learning is a Sequence Modeling Problem

Zero-sum Games:

A Unified Diversity Measure for Multiagent Reinforcement Learning

New RL environments:

Towards Human-Level Bimanual Dexterous Manipulation with Reinforcement Learning

MATE: Benchmarking Multi-Agent Reinforcement Learning in Distributed Target Coverage Control


Tutorial on Conference on Games 2022


Solving two-player zero-sum games through reinforcement learning

Part I

Part II


Two​ Invited Talks were given during the summer holidays


CSML China 08/22:

A continuum of solutions to cooperative MARL.

CCDM China 07/22:

Training a Population of Agents.


One paper got accepted at IROS 2022.

Fully Decentralized Model-based Policy Optimization for Networked Systems


We figured out how to do model-based MARL in networked systems.


One paper got accepted at IJCAI 2022.

1.  On the Convergence of Fictitious Play: A Decomposition Approach

     We extend the convergence guarantee for the well-known fictitious play method.


We open source two reinforcement learning projects:

1.  TorchOpt

     We develop an optimisation tool in Pytorch where meta-gradients can be                     computed easily.

     With TorchOpt, you can implement Meta-RL algorithms easily, try our code!

2.  BiDexHands

     We develop a RL/MARL environment for bimanual dexterous hands                           manipulations.

     BiDexhands are super fast, you can reach 40,000 FPS by only one GPU.


Two papers got accepted at ICLR 2022.


1.  Multiagent-Agent TRPO Methods

     We develop how to conduct trust-region updates in MARL settings.

     This is the SOTA algorithm in the cooperative MARL space, try our code!

     [English Blog]      [Chinese Blog]     [Code]

2.  LIGS: Learnable Intrinsic-Reward Generation Selection for Multi-          Agent Learning

     The paper addresses coordination improvement in the MARL setting by          learning intrinsic rewards that motivate the exploration and coordination.





Invited talk at DAI 2021 on the topic of Training A Population of Reinforcement Learning Agents.



Three papers get accepted at NeurIPS 2021:


          (see Blog here) 

We analysed the variance of gradient norm for multi-agent reinforcement learning and developed a minimal-variance policy gradient estimator.


We developed a rigorous way to generate diverse policies in population-based training and demonstrated impressive results on Google football. 


We show it is entirely possible to make AI learn to learn how to solve zero-sum games without even telling it what is a Nash equilibrium.


Invited talk at RLChina on the tutorial of Multi-Agent Learning.

Slides  Video (in Chinese)


Invited talk by 机器之心 on my recent work on how to deal with non-transitivity in two-player zero-sum games.

Slides  Video (in Chinese)


We opensource MALib: A bespoke high-performance framework for population-based multi-agent reinforcement learning.


Paper  Github


Two papers get accepted in ICML 2021.

Modelling Behavioural Diversity for Learning in Open-Ended Games. This paper studies how to measure and promote behavioural diversity in solving games in a mathematically rigorous way. It is awarded a long talk (top 3%) at ICML 2021.

Learning in Nonzero-Sum Stochastic Games with Potentials. This paper studies a generalised class of fully cooperative games, named stochastic potential games, and propose a MARL solution to find the Nash in such games. 


Check out my recent talk on the topic of:

A general framework for solving two-player zero-sum games. 


Update: Our paper wins the Best Paper Award at the Blue Sky Idea track!!!

One paper gets accepted in AAMAS 2021.

Diverse Auto-Curriculum is Critical for Successful Real-World Multiagent Learning Systems. I express some of my recent thoughts on why behavioural diversity in the policy space is an important factor for MARL techniques to be applied in real-world problems, outside purely video games. 


Check out my latest work on:

An Overview of Multi-Agent Reinforcement Learning from Game Theoretical Perspective.  I hope this work could offer a nice summary of game theory basics for MARL researches in addition to the deep RL hype :)


Update: SMARTS won the BEST paper award in CoRL 2020!

We release SMARTS: a multi-agent reinforcement learning enabled autonomous driving platform.

Multi-agent interaction is a fundamental aspect of autonomous driving in the real world. Today we are excited to introduce a dedicated platform: SMARTS, that supports Scalable Multi-Agent Reinforcement Learning Training for autonomous driving. With SMARTS, ML researchers can now evaluate their new algorithms in the self-driving scenarios, in addition to traditional video games. In turn, SMARTS can enrich the social vehicle behaviours and create increasingly more realistic and diverse interactions, powered by RL techniques, for autonomous driving researchers. Check our code on Github, and our paper at Conference on Robotic Learning 2020


One paper gets accepted at NIPS 2020 !

Replica-exchange Nos\'e-Hoover dynamics for Bayesian learning on large datasets. We introduce a new HMC sampler for large-scale Bayesian deep learning that suits multi-mode sampling and the noises from mini-batches can be absorbed by a special design of Nose-Hoover dynamics.


One paper gets accepted at CIKM 2020 !

Learning to infer user hidden states for online sequential advertising.


A lecture was given at RL China Summer School.

Advances of Multi-agent Learning in Gaming AI.


A talk was given at ISTBI, Fudan University.

Many-agent Reinforcement Learning.


One paper gets accepted at ICML 2020

Multi-agent Determinantal Q-learning. We introduce a new function approximator called Q-determinant point process for multi-agent reinforcement learning problems. It can help learn the Q-function factorisation with no needs for a priori structural constraints such as QMIX, VDN, etc.


One paper gets accepted at IJCAI 2020

Modelling Bounded Rationality in Multi-Agent Interactions by Generalized Recursive Reasoning. We use probabilisitic graphical model to describe the recursive reasoning process of "I believe you believe I believe..." in the multi-agent system.  


One paper gets accepted at AAMAS 2020

Alpha^Alpla-Rank: Practically Scaling Alpha-Rank through Stochastic Optimisation. Alpha-Rank is a replacement for Nash equilibrium for general-sum N-player game, importantly, its solution is P-complete. In this paper, we further enhance its tractability by several orders of magnitude by stochastic optimisation formulation. 

bottom of page