Sertan Girgin

ICLR Conference 2025 Conference Paper

BOND: Aligning LLMs with Best-of-N Distillation

• Pier Giuseppe Sessa
• Robert Dadashi
• Léonard Hussenot
• Johan Ferret
• Nino Vieillard
• Alexandre Ramé
• Bobak Shahriari
• Sarah Perrin

Reinforcement learning from human feedback (RLHF) is a key driver of quality and safety in state-of-the-art large language models. Yet, a surprisingly simple and strong inference-time strategy is Best-of-N sampling that selects the best generation among N candidates. In this paper, we propose Best-of-N Distillation (BOND), a novel RLHF algorithm that seeks to emulate Best-of-N but without its significant computational overhead at inference time. Specifically, BOND is a distribution matching algorithm that forces the distribution of generations from the policy to get closer to the Best-of-N distribution. We use the Jeffreys divergence (a linear combination of forward and backward KL) to balance between mode-covering and mode-seeking behavior, and derive an iterative formulation that utilizes a moving anchor for efficiency. We demonstrate the effectiveness of our approach and several design choices through experiments on abstractive summarization and Gemma models.

ICLR Conference 2025 Conference Paper

Diversity-Rewarded CFG Distillation

• Geoffrey Cideron
• Andrea Agostinelli
• Johan Ferret
• Sertan Girgin
• Romuald Elie
• Olivier Bachem
• Sarah Perrin
• Alexandre Ramé

Generative models are transforming creative domains such as music generation, with inference-time strategies like Classifier-Free Guidance (CFG) playing a crucial role. However, CFG doubles inference cost while limiting originality and diversity across generated contents. In this paper, we introduce diversity-rewarded CFG distillation, a novel finetuning procedure that distills the strengths of CFG while addressing its limitations. Our approach optimises two training objectives: (1) a distillation objective, encouraging the model alone (without CFG) to imitate the CFG-augmented predictions, and (2) an RL objective with a diversity reward, promoting the generation of diverse outputs for a given prompt. By finetuning, we learn model weights with the ability to generate high-quality and diverse outputs, without any inference overhead. This also unlocks the potential of weight-based model merging strategies: by interpolating between the weights of two models (the first focusing on quality, the second on diversity), we can control the quality-diversity trade-off at deployment time, and even further boost performance. We conduct extensive experiments on the MusicLM text-to-music generative model, where our approach surpasses CFG in terms of quality-diversity Pareto optimality. According to human evaluators, our finetuned-then-merged model generates samples with higher quality-diversity than the base model augmented with CFG. Explore our generations at https://musicdiversity.github.io/.

ICML Conference 2024 Conference Paper

MusicRL: Aligning Music Generation to Human Preferences

• Geoffrey Cideron
• Sertan Girgin
• Mauro Verzetti
• Damien Vincent
• Matej Kastelic
• Zalán Borsos
• Brian McWilliams
• Victor Ungureanu

We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e. g. a caption such as “upbeat workout music” can map to a retro guitar solo or a technopop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM model of discrete audio tokens finetuned with reinforcement learning to maximize sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300, 000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models. Samples can be found at google-research. github. io/seanet/musiclm/rlhf/.

ICML Conference 2024 Conference Paper

Nash Learning from Human Feedback

• Rémi Munos
• Michal Valko
• Daniele Calandriello
• Mohammad Gheshlaghi Azar
• Mark Rowland 0001
• Daniel Guo 0001
• Yunhao Tang
• Matthieu Geist

Reinforcement learning from human feedback (RLHF) has emerged as the main paradigm for aligning large language models (LLMs) with human preferences. Traditionally, RLHF involves the initial step of learning a reward model from pairwise human feedback, i. e. , expressed as preferences between pairs of text generations. Subsequently, the LLM’s policy is fine-tuned to maximize the reward through a reinforcement learning algorithm. In this study, we introduce an alternative pipeline for the fine-tuning of LLMs using pairwise human feedback. Our approach entails the initial learning of a pairwise preference model, which is conditioned on two inputs (instead of a single input in the case of a reward model) given a prompt, followed by the pursuit of a policy that consistently generates responses preferred over those generated by any competing policy, thus defining the Nash equilibrium of this preference model. We term this approach Nash learning from human feedback (NLHF). In the context of a tabular policy representation, we present a novel algorithmic solution, Nash-MD, founded on the principles of mirror descent. This algorithm produces a sequence of policies, with the last iteration converging to the regularized Nash equilibrium. Additionally, we explore parametric representations of policies and introduce gradient descent algorithms for deep-learning architectures. We illustrate the effectiveness of our approach by presenting experimental results on a text summarization task. We believe NLHF offers a compelling avenue for fine-tuning LLMs and enhancing the alignment of LLMs with human preferences.

ICML Conference 2022 Conference Paper

Continuous Control with Action Quantization from Demonstrations

• Robert Dadashi
• Léonard Hussenot
• Damien Vincent
• Sertan Girgin
• Anton Raichuk
• Matthieu Geist
• Olivier Pietquin

In this paper, we propose a novel Reinforcement Learning (RL) framework for problems with continuous action spaces: Action Quantization from Demonstrations (AQuaDem). The proposed approach consists in learning a discretization of continuous action spaces from human demonstrations. This discretization returns a set of plausible actions (in light of the demonstrations) for each input state, thus capturing the priors of the demonstrator and their multimodal behavior. By discretizing the action space, any discrete action deep RL technique can be readily applied to the continuous control problem. Experiments show that the proposed approach outperforms state-of-the-art methods such as SAC in the RL setup, and GAIL in the Imitation Learning setup. We provide a website with interactive videos: https: //google-research. github. io/aquadem/ and make the code available: https: //github. com/google-research/google-research/tree/master/aquadem.

EWRL Workshop 2022 Workshop Paper

Continuous Control with Action Quantization from Demonstrations

• Robert Dadashi
• Léonard Hussenot
• Damien Vincent
• Sertan Girgin
• Anton Raichuk
• Matthieu Geist
• Olivier Pietquin

In this paper, we propose a novel Reinforcement Learning (RL) framework for problems with continuous action spaces: Action Quantization from Demonstrations (AQuaDem). The proposed approach consists in learning a discretization of continuous action spaces from human demonstrations. This discretization returns a set of plausible actions (in light of the demonstrations) for each input state, thus capturing the priors of the demonstrator and their multimodal behavior. By discretizing the action space, any discrete action deep RL technique can be readily applied to the continuous control problem. Experiments show that the proposed approach outperforms state-of-the-art methods such as SAC in the RL setup, and GAIL in the Imitation Learning setup.

EWRL Workshop 2022 Workshop Paper

Get Back Here: Robust Imitation by Return-to-Distribution Planning

• Geoffrey Cideron
• Olivier Pietquin
• Robert Dadashi
• Gabriel Dulac-Arnold
• Baruch Tabanpour
• Matthieu Geist
• Léonard Hussenot
• Sebastian Curi

We consider the Imitation Learning (IL) setup where expert data are not collected on the actual deployment environment but on a different version. To address the resulting distribution shift, we combine behavior cloning (BC) with a planner that is tasked to bring the agent back to states visited by the expert whenever the agent deviates from the demonstration distribution. The resulting algorithm, POIR, can be trained offline, and leverages online interactions to fine-tune its planner to improve performance over time. We test POIR on a variety of human-generated manipulation demonstrations and show robustness of the learned policy to different initial state distributions and different dynamics.

EWRL Workshop 2022 Workshop Paper

Scalable Deep Reinforcement Learning Algorithms for Mean Field Games

• Mathieu Lauriere
• Sarah Perrin
• Sertan Girgin
• Paul Muller
• Ayush Jain
• Théophile Cabannes
• Georgios Piliouras
• Julien Perolat

Mean Field Games (MFGs) have been introduced to efficiently approximate games with very large populations of strategic agents. Recently, the question of learning equilibria in MFGs has gained momentum, particularly using model-free reinforcement learning (RL) methods. One limiting factor to further scale up using RL is that existing algorithms to solve MFGs require the mixing of approximated quantities such as strategies or q-values. This is non-trivial in the case of non-linear function approximation that enjoy good generalization properties, e.g. neural networks. We propose two methods to address this shortcoming. The first one learns a mixed strategy from distillation of historical data into a neural network and is applied to the Fictitious Play algorithm. The second one is an online mixing method based on regularization that does not require memorizing historical data or previous estimates. It is used to extend Online Mirror Descent. We demonstrate numerically that these methods efficiently enable the use of Deep RL algorithms to solve various MFGs. In addition, we show that these methods outperform SotA baselines from the literature.

ICML Conference 2022 Conference Paper

Scalable Deep Reinforcement Learning Algorithms for Mean Field Games

• Mathieu Laurière
• Sarah Perrin
• Sertan Girgin
• Paul Muller
• Ayush Jain
• Theophile Cabannes
• Georgios Piliouras
• Julien Pérolat

Mean Field Games (MFGs) have been introduced to efficiently approximate games with very large populations of strategic agents. Recently, the question of learning equilibria in MFGs has gained momentum, particularly using model-free reinforcement learning (RL) methods. One limiting factor to further scale up using RL is that existing algorithms to solve MFGs require the mixing of approximated quantities such as strategies or $q$-values. This is far from being trivial in the case of non-linear function approximation that enjoy good generalization properties, e. g. neural networks. We propose two methods to address this shortcoming. The first one learns a mixed strategy from distillation of historical data into a neural network and is applied to the Fictitious Play algorithm. The second one is an online mixing method based on regularization that does not require memorizing historical data or previous estimates. It is used to extend Online Mirror Descent. We demonstrate numerically that these methods efficiently enable the use of Deep RL algorithms to solve various MFGs. In addition, we show that these methods outperform SotA baselines from the literature.

AAMAS Conference 2022 Conference Paper

Solving N-Player Dynamic Routing Games with Congestion: A Mean-Field Approach

• Theophile Cabannes
• Mathieu Laurière
• Julien Perolat
• Raphael Marinier
• Sertan Girgin
• Sarah Perrin
• Olivier Pietquin
• Alexandre M. Bayen

The recent emergence of navigational tools has changed traffic patterns and has now enabled new types of congestion-aware routing control like dynamic road pricing. Using the fundamental diagram of traffic flows – applied in macroscopic and mesoscopic traffic modeling – the article introduces a new 𝑁-player dynamic routing game with explicit congestion dynamics. The model is well-posed and can reproduce heterogeneous departure times and congestion spill back phenomena. However, as Nash equilibrium computations are PPAD-complete, solving the game becomes intractable for large but realistic numbers of vehicles 𝑁. Therefore, the corresponding mean field game is also introduced. Experiments were performed on several classical benchmark networks of the traffic community: the Pigou, Braess, and Sioux Falls networks with heterogeneous origin, destination and departure time tuples. The Pigou and the Braess examples reveal that the mean field approximation is generally very accurate and computationally efficient as soon as the number of vehicles exceeds a few dozen. On the Sioux Falls network (76 links, 100 time steps), this approach enables learning traffic dynamics with more than 14, 000 vehicles.

NeurIPS Conference 2021 Conference Paper

Brax - A Differentiable Physics Engine for Large Scale Rigid Body Simulation

• Daniel Freeman
• Erik Frey
• Anton Raichuk
• Sertan Girgin
• Igor Mordatch
• Olivier Bachem

We present Brax, an open source library for \textbf{r}igid \textbf{b}ody simulation with a focus on performance and parallelism on accelerators, written in JAX. We present results on a suite of tasks inspired by the existing reinforcement learning literature, but remade in our engine. Additionally, we provide reimplementations of PPO, SAC, ES, and direct policy optimization in JAX that compile alongside our environments, allowing the learning algorithm and the environment processing to occur on the same device, and to scale seamlessly on accelerators. Finally, we include notebooks that facilitate training of performant policies on common MuJoCo-like tasks in minutes.

ICML Conference 2021 Conference Paper

Hyperparameter Selection for Imitation Learning

• Léonard Hussenot
• Marcin Andrychowicz
• Damien Vincent
• Robert Dadashi
• Anton Raichuk
• Sabela Ramos
• Nikola Momchev
• Sertan Girgin

We address the issue of tuning hyperparameters (HPs) for imitation learning algorithms in the context of continuous-control, when the underlying reward function of the demonstrating expert cannot be observed at any time. The vast literature in imitation learning mostly considers this reward function to be available for HP selection, but this is not a realistic setting. Indeed, would this reward function be available, it could then directly be used for policy training and imitation would not be necessary. To tackle this mostly ignored problem, we propose a number of possible proxies to the external reward. We evaluate them in an extensive empirical study (more than 10’000 agents across 9 environments) and make practical recommendations for selecting HPs. Our results show that while imitation learning algorithms are sensitive to HP choices, it is often possible to select good enough HPs through a proxy to the reward function.

NeurIPS Conference 2021 Conference Paper

What Matters for Adversarial Imitation Learning?

• Manu Orsini
• Anton Raichuk
• Leonard Hussenot
• Damien Vincent
• Robert Dadashi
• Sertan Girgin
• Matthieu Geist
• Olivier Bachem

Adversarial imitation learning has become a popular framework for imitation in continuous control. Over the years, several variations of its components were proposed to enhance the performance of the learned policies as well as the sample complexity of the algorithm. In practice, these choices are rarely tested all together in rigorous empirical studies. It is therefore difficult to discuss and understand what choices, among the high-level algorithmic options as well as low-level implementation details, matter. To tackle this issue, we implement more than 50 of these choices in a generic adversarial imitation learning frameworkand investigate their impacts in a large-scale study (>500k trained agents) with both synthetic and human-generated demonstrations. We analyze the key results and highlight the most surprising findings.

ICLR Conference 2021 Conference Paper

What Matters for On-Policy Deep Actor-Critic Methods? A Large-Scale Study

• Marcin Andrychowicz
• Anton Raichuk
• Piotr Stanczyk
• Manu Orsini
• Sertan Girgin
• Raphaël Marinier
• Léonard Hussenot
• Matthieu Geist

In recent years, reinforcement learning (RL) has been successfully applied to many different continuous control tasks. While RL algorithms are often conceptually simple, their state-of-the-art implementations take numerous low- and high-level design decisions that strongly affect the performance of the resulting agents. Those choices are usually not extensively discussed in the literature, leading to discrepancy between published descriptions of algorithms and their implementations. This makes it hard to attribute progress in RL and slows down overall progress [Engstrom'20]. As a step towards filling that gap, we implement >50 such ``"choices" in a unified on-policy deep actor-critic framework, allowing us to investigate their impact in a large-scale empirical study. We train over 250'000 agents in five continuous control environments of different complexity and provide insights and practical recommendations for the training of on-policy deep actor-critic RL agents.

EWRL Workshop 2008 Conference Paper

Basis Expansion in Natural Actor Critic Methods

• Sertan Girgin
• Philippe Preux

Abstract In reinforcement learning, the aim of the agent is to find a policy that maximizes its expected return. Policy gradient methods try to accomplish this goal by directly approximating the policy using a parametric function approximator; the expected return of the current policy is estimated and its parameters are updated by steepest ascent in the direction of the gradient of the expected return with respect to the policy parameters. In general, the policy is defined in terms of a set of basis functions that capture important features of the problem. Since the quality of the resulting policies directly depend on the set of basis functions, and defining them gets harder as the complexity of the problem increases, it is important to be able to find them automatically. In this paper, we propose a new approach which uses cascade-correlation learning architecture for automatically constructing a set of basis functions within the context of Natural Actor-Critic (NAC) algorithms. Such basis functions allow more complex policies be represented, and consequently improve the performance of the resulting policies. We also present the effectiveness of the method empirically.

IJCAI Conference 2007 Conference Paper

• Sertan Girgin
• Faruk Polat
• Reda Alhajj

This paper employs state similarity to improve reinforcement learning performance. This is achieved by first identifying states with similar sub-policies. Then, a tree is constructed to be used for locating common action sequences of states as derived from possible optimal policies. Such sequences are utilized for defining a similarity function between states, which is essential for reflecting updates on the action-value function of a state onto all similar states. As a result, the experience acquired during learning can be applied to a broader context. Effectiveness of the method is demonstrated empirically.

ECAI Conference 2006 Conference Paper

Learning by Automatic Option Discovery from Conditionally Terminating Sequences

• Sertan Girgin
• Faruk Polat
• Reda Alhajj

This paper proposes a novel approach to discover options in the form of conditionally terminating sequences, and shows how they can be integrated into reinforcement learning framework to improve the learning performance. The method utilizes stored histories of possible optimal policies and constructs a specialized tree structure online in order to identify action sequences which are used frequently together with states that are visited during the execution of such sequences. The tree is then used to implicitly run corresponding options. Effectiveness of the method is demonstrated empirically.