Yunhao (Robin) Tang

I am a research scientist at DeepMind London. I am a core contributor to the Gemini post-training and I research the science and engineering of deep reinforcement learning. Previously, I was a two-time intern at DeepMind Paris hosted by Remi Munos. I obtained my PhD at Columbia University in New York City.

Email  /  CV (updated Oct, 2021)  /  Google Scholar  /  Twitter

profile photo
News
  • [7/2024] Great to be part of the project led by my excellent collaborators to benchmark scalable-oversight protocols. Check it out here.
  • [5/2024] We have a new paper that hypothesis-tests the importance of on-policy sampling in language model alignment, check it out here!
  • [5/2024] Four papers are accepted at ICML 2024. Thank you for the heavy lifting to my coauthors.
  • [3/2024] Gemini 1.5 is announced. Check out the tech report here.
  • [12/2023] Gemini is launched. Great to be a core contributor to Gemini, the most powerful multi-modal large language model developed by Google DeepMind. Check out the tech report here.
Research

My recent work focuses on the understanding and developments of deep reinforcement learning algorithms and systems, spanning the following non-exhaustive list of topics

  • Learning from human feedback
  • Representation learning
  • Distributional reinforcement learning
  • Search and credit assignment
  • Off-policy reinforcement learning
  • Stochastic gradient estimation

See here for the full list of publications.

Selected Publications
Understanding the Performance Gap between Online and Offline Alignment Algorithms
Yunhao Tang, Daniel Guo, Zeyu Zheng, Daniele Calandriello, Yuan Cao, Eugene Tarassov, Remi Munos, Bernardo Avila Pires, Michal Valko, Yong Cheng, Will Dabney
Arxiv,

Is online RL really necessary for AI alignment, or do offline algorithms suffice? The answer seems to be yes according to our careful ablations.

On scalable oversight with weak LLMs judging strong LLMs
Zachary Kenton*, Noah Y. Siegel*, Janos Kramar, Jonah Brown-Cohen, Samuel Albanie, Jannis Bulian, Rishabh Agarwal, David Lindner, Yunhao Tang, Noah D. Goodman, Rohin Shah
Arxiv,

We have benchmarked important existing scalable-oversight protocols in a comprehensive suite of QA tasks, opening the path for further future investigation.

Offline Regularised Reinforcement Learning for Large Language Models Alignment
Pierre Harvey Richemond, Yunhao Tang, Daniel Guo, Daniele Calandriello, Mohammad Gheshlaghi Azar, Rafael Rafailov, Bernardo Avila Pires, Eugene Tarassov, Lucas Spangher, Will Ellsworth, Aliaksei Severyn, Jonathan Mallinson, Lior Shani, Gil Shamir, Rishabh Joshi, Tianqi Liu, Remi Munos, Bilal Piot
Arxiv,

When human feedback is pointwise rather than pairwise, we propose direct reward optimization (DRO) as the alignment algorithm.

Human Alignment of Large Language Models Through Online Preference Optimization
Daniele Calandriello, Daniel Guo, Remi Munos, Mark Rowland, Yunhao Tang, Bernardo Avila Pires, Pierre Harvey Richemond, Charline Le Lan, Michal Valko, Tianqi Liu, Rishabh Joshi, Zeyu Zheng, Bilal Piot
Arxiv, ICML 2024

Online preference optimization as an alignment technique turns out to be intimately related to Nash equilibrium, besides being a competitive algorithm for RLHF.

Generalized Preference Optimization: A Unified Approach to Offline Alignment
Yunhao Tang, Daniel Zhaohan Guo, Zeyu Zheng, Daniele Calandriello, Remi Munos, Mark Rowland, Pierre Harvey Richmond, Michal Valko, Bernardo Avila Pires, Bilal Piot
Arxiv, ICML 2024

GPO unifies alignment algorithms such as DPO, IPO and SLiC as special cases. The insight, interestingly, is based on classic literature on convex losses for binary classification. At the end of the day, all algorithmic variants have similar performance-regularization trade-off though their natural strengths of regularization differ.

Gemini: A Family of Highly Capable Multimodal Models
Gemini team, Google DeepMind.
Tech report Arxiv,

One of the most powerful multi-modal large language models thus far in the world.

Nash Learning from Human Feedback
Remi Munos*, Michal Valko*, Daniele Calandriello*, Mohammad Gheshlaghi Azar*, Mark Rowland*, Daniel Guo*, Yunhao Tang*, Matthieu Geist*, Thomas Mesnard, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J. Mankowitz, Doina Precup and Bilal Piot*
Arxiv, ICML 2024

In aligning large language models, we search for Nash Equilibrium naturally defined via the pairwise human feedback. This approach is more general purpose, imposes fewer assumptions on reward modeling, and performs better than canonical RLHF.

Off-policy Distributional Q(lambda): Distributional RL without Importance Sampling
Yunhao Tang, Mark Rowland, Remi Munos, Bernardo Avila Pires, Will Dabney
Arxiv,

We introduce another addition to the family of off-policy distributional RL algorithms, importantly, without the need for importance sampling.

A Distributional Analogue to the Successor Representation
Harley Wiltzer*, Jesse Farebrother*, Arthur Greton, Yunhao Tang, André Barreto, Will Dabney, Marc G. Bellemare, Mark Rowland
Arxiv, ICML 2024

We shed light on what distributional equivalence of successor representations look like, and the algorithmic applications arising from the insights.

VA-learning as a more efficient alternative to Q-learning
Yunhao Tang, Remi Munos, Mark Rowland, Michal Valko
Arxiv, ICML 2023

We propose VA-learning as a more sample efficient alternative to Q-learning. The sample efficiency stems from the value sharing between different actions. Intriguingly, VA-learning closely relates to dueling architecture.

DoMo-AC: Doubly Multi-step Off-policy Actor-Critic Algorithm
Yunhao Tang*, Tadashi Kozuno*, Mark Rowland, Anna Harutyunyan, Remi Munos, Bernardo Avila Pires, Michal Valko
Arxiv, ICML 2023

We design an off-policy actor-critic algorithm based on multi-step policy improvement and policy evaluation. This algorithm improves state-of-the-art IMPALA baseline.

Towards a better understanding of representation dynamics under TD-learning
Yunhao Tang, Remi Munos
Arxiv, ICML 2023

We provide a characterization on how TD-learning learns representations, relating random reward based TD-learning with spectral decomposition of the transition matrix.

Quantile Credit Assignment
Thomas Mesnard, Wenqi Chen, Alaa Saade, Yunhao Tang, Mark Rowland, Theophane Weber, Clare Lyle, Audrunas Gruslys, Michal Valko, Will Dabney, Georg Ostrovski, Eric Moulines, Remi Munos
Arxiv, ICML 2023, Oral

Efficient credit assignment should account for external factors outside of agent's control, or more informally, the level of luck. We formalize such intuitions into quantile credit assignment.

The Statistical Benefits of Quantile Temporal-Difference Learning for Value Estimation
Mark Rowland, Yunhao Tang, Clare Lyle, Remi Munos, Marc G. Bellemare, Will Dabney
Arxiv, ICML 2023

We show that in certain cases quantile TD outperforms TD in mean value prediction. This hints at a general potential of distributional RL to outperform mean-based RL at its own game.

Understanding Self-Predictive Learning for Reinforcement Learning
Yunhao Tang, Zhaohan Daniel Guo, Pierre Harvey Richemond, Bernardo Avila Pires, Yash Chandak, Remi Munos, Mark Rowland, Mohammad Gheshlaghi Azar, Charline Le Lan, Clare Lyle, Andras Gyorgy, Shantanu Thakoor, Will Dabney, Bilal Piot, Daniele Calandriello, Michal Valko
Arxiv, ICML 2023

Self-predictive learning is a popular representation learning algorithm in RL, which learns a latent representation by predicting (bootstrapping) its own future latents. Intuitively, the algorithm should not work as it can collapse to trivial solutions. We identify algorithmic components to prevent the collapse and show that self-preditive learning is related to gradient-based spectral decomposition of the transition dynamics.

An Analysis of Quantile Temporal-Difference Learning
Mark Rowland, Remi Munos, Mohammad Gheshlaghi Azar, Yunhao Tang, Georg Ostrovski, Anna Harutyunyan, Karl Tuyls, Marc G. Bellemare, Will Dabney
JMLR
Arxiv,

We provide a first proof of the convergence of quantile TD-learning, a distributional RL algorithm that drives multiple recent empirical breakthroughs.

The Nature of Temporal Difference Errors in Multi-step Distributional Reinforcement Learning
Yunhao Tang, Mark Rowland, Remi Munos, Bernardo Avila Pires, Will Dabney, Marc G. Bellemare
Arxiv, NeurIPS 2022

We identify a few intriguing and fundamental differences between value-based TD-learning and distributional TD-learning.

BYOL-Explore: Exploration by Bootstrapped Prediction
Zhaohan Daniel Guo*, Shantanu Thakoor*, Miruna Pislar*, Bernardo Avila Pires*, Florent Altche*, Corentin Tallec*, Alaa Saade, Daniele Calandriello, Jean-Bastien Grill, Yunhao Tang, Michal Valko, Remi Munos, Mohammad Gheshlaghi Azar*, Bilal Piot*
Arxiv, NeurIPS 2022

We find that self-prediction loss is a surprisingly useful signal for exploration in extremely challenging deep RL domains. Our method: BYOL-explore, partially cracks a wide range of extremely hard exploration problems much more efficiently than prior methods.

Biased Gradient Estimate with Drastic Variance Reduction for Meta Reinforcement Learning
Yunhao Tang
International Conference on Machine Learning (ICML), Baltimore, USA, 2022
arXiv

We find that certain deliberate bias in gradient estimators could significantly reduce variance for meta RL.

Unifying Gradient Estimators for Meta-Reinforcement Learning via Off-Policy Evaluation
Yunhao Tang*, Tadashi Kozuno*, Mark Rowland, Remi Munos, Michal Valko
Neural Information Processing Systems (NeurIPS), Virtual, 2021
arXiv Code

How to estimate high-order derivatives of value functions? We propose a unifying framework with off-policy evaluation. Direct differentiations of off-policy estimates produce estimates to high-order derivatives of value functions, and instantiate many prior methods as special cases.

Revisiting Peng's Q($\lambda$) for Modern Reinforcement Learning
Tadashi Kozuno*, Yunhao Tang*, Mark Rowland, Remi Munos, Steven Kapturowski, Will Dabney, Michal Valko, David Abel
International Conference on Machine Learning (ICML), Virtual, 2021
ArXiv Code

Uncorrected multi-step updates such as n-step Q-learning are ubiquitous in modern deep RL practices. We revisit Peng's Q($\lambda$), a classic uncorrected multi-step variant. Our analysis sheds light on why uncorrected updates should work in practice. The empirical result also suggests significant gains on benchmark tasks.

Hindsight Expectation Maximization for Goal-conditioned Reinforcement Learning
Yunhao Tang, Alp Kucukelbir
International Conference on Artificial Intelligence and Statistics (AISTATS), Virtual, 2021
paper / arXiv

We propose Hindsight Expectation Maximization (hEM), an EM algorithm for goal-conditioned RL problem which combines supervised learning through the M-step and hindsight goal sampling through the E-step. We also make an intimate connection between hindsight replay and importance sampling for rare event simulations.

Self-Imitation Learning via Generalized Lower Bound Q-learning
Yunhao Tang
Neural Information Processing Systems (NeurIPS), Vancouver, Canada, 2020
paper / arXiv

Why is self-imitation learning efficient? We shed light on its connections to n-step Q-learning and show that part of its gains might be attributed to trade-offs in RL operators. We also propose a n-step extension of self-imitation learning which incorporates the strengths of both n-step updates and lower-bound learning.

Monte-Carlo Tree Search as Regularized Policy Optimization
Jean-Bastien Grill*, Florent Altche*, Yunhao Tang*, Thomas Hubert, Michal Valko, Ioannis Antonoglou, Remi Munos
International Conference on Machine Learning (ICML), Vienna, Austria, 2020
paper / arXiv / video

We establish an interpretation of MCTS as policy optimization. This interpretation leads to algorithmic variants which naturally improve over MCTS-based baselines such as AlphaZero and MuZero.

Taylor Expansion Policy Optimization
Yunhao Tang, Michal Valko, Remi Munos
International Conference on Machine Learning (ICML), Vienna, Austria, 2020
paper / arXiv / video / media 1 / media 2

We estabilish the intimate connections between trust region policy search and off-policy evaluation. The new algorithm TayPO generalizes policy optimization objectives to high-order extentions which leads to gains on large-scale distributed agents.

Reinforcement Learning for Integer Programming: Learning to Cut
Yunhao Tang, Shipra Agrawal, Yuri Faenza
International Conference on Machine Learning (ICML), Vienna, Austria, 2020
paper / arXiv / video

We formulate cutting plane algorithms as a sequential decision making problem for generic integer pgroamming. The cutting plane agent learned via RL improves over human-designed heuristics and benfits downstream applications such as branch-and-cut.

The source code of the website is from here.