Glen Berseth

Biographie

Glen Berseth est professeur agrégé au Département d'informatique et de recherche opérationnelle (DIRO) de l'Université de Montréal, membre académique principal de Mila – Institut québécois d'intelligence artificielle, détenteur d’une chaire en IA Canada-CIFAR et codirecteur du Laboratoire de robotique et d’IA intégrative de Montréal (REAL). Il a été chercheur postdoctoral à Berkeley Artificial Intelligence Research (BAIR), où il a travaillé avec Sergey Levine. Ses recherches portent sur la résolution de problèmes de prise de décision séquentielle (planification) pour les systèmes d'apprentissage autonomes du monde réel (robots). Elles ont couvert les domaines de la collaboration humain-robot, du renforcement, ainsi que de l'apprentissage continu, multiagent et hiérarchique et du méta-apprentissage. Glen Berseth a fait paraître des articles dans les meilleures publications des domaines de la robotique, de l'apprentissage automatique et de l'animation informatique. Il donne également un cours sur l'apprentissage des robots à l'Université de Montréal et à Mila, couvrant les recherches les plus récentes sur les techniques d'apprentissage automatique pour la création de robots généralistes.

Étudiants actuels

Doctorat - UdeM

Doctorat - UdeM

Maîtrise recherche - UdeM

Roger Creus-Castanyer

Doctorat - UdeM

Co-superviseur⋅e :

Doctorat - McGill

Superviseur⋅e principal⋅e :

Hsiu-Chin Lin

Léa Demeule

Doctorat - UdeM

Co-superviseur⋅e :

Doctorat - UdeM

Superviseur⋅e principal⋅e :

Liam Paull

Kumaraditya Gupta

Doctorat

Superviseur⋅e principal⋅e :

Collaborateur·rice de recherche - UdeM

Adriana Knatchbull-Hugessen

Doctorat - UdeM

Artur Kuramshin

Maîtrise recherche - UdeM

Daniel Lawson

Doctorat - UdeM

Co-superviseur⋅e :

Postdoctorat - UdeM

Co-superviseur⋅e :

Maîtrise recherche - UdeM

Stagiaire de recherche - UdeM

Esra'a Saleh

Doctorat - UdeM

Co-superviseur⋅e :

Doctorat - UdeM

Co-superviseur⋅e :

Marc Gendron-Bellemare

Collaborateur·rice de recherche

L'apprentissage par renforcement en temps réel

Siddarth Venkatraman

Doctorat - UdeM

Co-superviseur⋅e :

Albert Zhan

Doctorat - UdeM

Billets de blogue

Deux robots dans une cuisine, en train de préparer le dîner. L'un coupe les légumes et l'autre fait une omelette.

20 juin 2025

par

Ivan Anokhin

Matthew Riemer

Rishav Rishav

Gopeshh Subbaraj

Glen Berseth

Lire l'article

Fully Autonomous Real-World Reinforcement Learning with Applications to Mobile Manipulation

15 février 2023

Apprentissage par renforcement entièrement autonome dans le monde réel avec des applications à la manipulation mobile

par

Jędrzej Orbik

Charles Sun

Coline Devin

Glen Berseth

Lire l'article

Publications

A Comedy of Estimators: On KL Regularization in RL Training of LLMs

Vedant Shah

Johan Obando-Ceron

Vineet Jain

Brian Bartoldson

Bhavya Kailkhura

Nikolay Malkin

The reasoning performance of large language models (LLMs) can be substantially improved by training them with reinforcement learning (RL). T… (voir plus)he RL objective for LLM training involves a regularization term, which is the reverse Kullback-Leibler (KL) divergence between the trained policy and the reference policy. Since computing the KL divergence exactly is intractable, various estimators are used in practice to estimate it from on-policy samples. Despite its wide adoption, including in several open-source libraries, there is no systematic study analyzing the numerous ways of incorporating KL estimators in the objective and their effect on the downstream performance of RL-trained models. Recent works show that prevailing practices for incorporating KL regularization do not provide correct gradients for stated objectives, creating a discrepancy between the objective and its implementation. In this paper, we further analyze these practices and study the gradients of several estimators configurations, revealing how design choices shape gradient bias. We substantiate these findings with empirical observations by RL fine-tuning \texttt{Qwen2.5-7B}, \texttt{Llama-3.1-8B-Instruct} and \texttt{Qwen3-4B-Instruct-2507} with different configurations and evaluating their performance on both in- and out-of-distribution tasks. Through our analysis, we observe that, in on-policy settings: (1) estimator configurations with biased gradients can result in training instabilities; and (2) using estimator configurations resulting in unbiased gradients leads to better performance on in-domain as well as out-of-domain tasks. We also investigate the performance resulting from different KL configurations in off-policy settings and observe that KL regularization can help stabilize off-policy RL training resulting from asynchronous setups.

2025-12-25

ArXiv (prépublication)

arxiv.org

RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies

Pranav Atreya

Karl Pertsch

Tony Lee

Moo Jin Kim

Arhan Jain

Artur Kuramshin

Cyrus Neary

Edward S. Hu

Kanav Arora

Kirsty Ellis

Luca Macesanu

Matthew Leonard

Meedeum Cho

Özgür Aslan

Shivin Dass

Tony Wang

Xingfang Yuan

Abhishek Gupta

Dinesh Jayaraman

Glen Berseth … (voir 6 de plus)

Kostas Daniilidis

Roberto Martín-Martín

Youngwoon Lee

Percy Liang

Chelsea Finn

Sergey Levine

2025-10-06

Proceedings of The 8th Conference on Robot Learning (publié)

proceedings.mlr.press

Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models

Johan Samir Obando Ceron

Yoshua Bengio

Brian R. Bartoldson

Bhavya Kailkhura

Guillaume Lajoie

Nikolay Malkin

Moksh J. Jain

Test-time scaling methods improve the capabilities of large language models (LLMs) by increasing the amount of compute used during inference… (voir plus) to make a prediction. Inference-time compute can be scaled in parallel by choosing among multiple independent solutions or sequentially through self-refinement. We propose Recursive Self-Aggregation (RSA), a test-time scaling method inspired by evolutionary methods that combines the benefits of both parallel and sequential scaling. Each step of RSA refines a population of candidate reasoning chains through aggregation of subsets to yield a population of improved solutions, which are then used as the candidate pool for the next iteration. RSA exploits the rich information embedded in the reasoning chains -- not just the final answers -- and enables bootstrapping from partially correct intermediate steps within different chains of thought. Empirically, RSA delivers substantial performance gains with increasing compute budgets across diverse tasks, model families and sizes. Notably, RSA enables Qwen3-4B-Instruct-2507 to achieve competitive performance with larger reasoning models, including DeepSeek-R1 and o3-mini (high), while outperforming purely parallel and sequential scaling strategies across AIME-25, HMMT-25, Reasoning Gym, LiveCodeBench-v6, and SuperGPQA. We further demonstrate that training the model to combine solutions via a novel aggregation-aware reinforcement learning approach yields significant performance gains. Code available at https://github.com/HyperPotatoNeo/RSA.

2025-09-29

ArXiv (prépublication)

arxiv.org

Concept-based Steering of Large Language Models for Conditional Molecular Generation

Yang Zhang

Modern LLMs, with their internet-scale pretraining and advanced human-level capabilities across specialized tasks, have demonstrated promisi… (voir plus)ng performance in molecular discovery using existing text-based molecular representations, such as SMILES and SELFIES. However, generating valid, unique, and high-fidelity molecules while precisely controlling for multiple properties simultaneously remains challenging. While prior works demonstrated success by fine-tuning language models on a novel corpus of molecules with property-conditioned tags, real-world applications require generating molecules from diverse property distributions, previously unseen in the training data. To this end, we present Concept-based Activation STeering (CAST), the first approach to apply activation steering to directly edit a model's internal representation for conditional molecular generation. CAST offers a lightweight, flexible alternative to fine-tuning by computing property-conditioned steering vectors via a concept network that does not require retraining the LLM. Through extensive experiments on datasets such as Therapeutics Data Commons, we show that CAST consistently outperforms existing methods on both in-distribution and out-of-distribution conditional generation tasks. We also conduct comprehensive ablation studies to highlight the extent of control our concept-guided steering provides on the molecules generated by the LLM.

2025-09-19

NeurIPS.cc/2025/Workshop/AI4Mat (poster)

Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning

Roger Creus Castanyer

Johan Obando-Ceron

Scaling deep reinforcement learning networks is challenging and often results in degraded performance, yet the root causes of this failure m… (voir plus)ode remain poorly understood. Several recent works have proposed mechanisms to address this, but they are often complex and fail to highlight the causes underlying this difficulty. In this work, we conduct a series of empirical analyses which suggest that the combination of non-stationarity with gradient pathologies, due to suboptimal architectural choices, underlie the challenges of scale. We propose a series of direct interventions that stabilize gradient flow, enabling robust performance across a range of network depths and widths. Our interventions are simple to implement and compatible with well-established algorithms, and result in an effective mechanism that enables strong performance even at large scales. We validate our findings on a variety of agents and suites of environments.

2025-09-17

NeurIPS.cc/2025/Conference (spotlight)

Task Robustness via Re-Labelling Vision-Action Robot Data

Artur Kuramshin

Özgür Aslan

Cyrus Neary

2025-09-05

robot-learning.org/CoRL/2025/Workshop/Robot_Data (publié)

Adaptive Resolution Residual Networks — Generalizing Across Resolutions Easily and Efficiently

Léa Demeule

Mahtab Sandhu

The majority of signal data captured in the real world uses numerous sensors with different resolutions. In practice, most deep learning arc… (voir plus)hitectures are fixed-resolution; they consider a single resolution at training and inference time. This is convenient to implement but fails to fully take advantage of the diverse signal data that exists. In contrast, other deep learning architectures are adaptive-resolution; they directly allow various resolutions to be processed at training and inference time. This provides computational adaptivity but either sacrifices robustness or compatibility with mainstream layers, which hinders their use. In this work, we introduce Adaptive Resolution Residual Networks (ARRNs) to surpass this tradeoff. We construct ARRNs from Laplacian residuals, which serve as generic adaptive-resolution adapters for fixed-resolution layers. We use smoothing filters within Laplacian residuals to linearly separate input signals over a series of resolution steps. We can thereby skip Laplacian residuals to cast high-resolution ARRNs into low-resolution ARRNs that are computationally cheaper yet numerically identical over low-resolution signals. We guarantee this result when Laplacian residuals are implemented with perfect smoothing kernels. We complement this novel component with Laplacian dropout, which randomly omits Laplacian residuals during training. This regularizes for robustness to a distribution of lower resolutions. This also regularizes for numerical errors that may occur when Laplacian residuals are implemented with approximate smoothing kernels. We provide a solid grounding for the advantageous properties of ARRNs through a theoretical analysis based on neural operators, and empirically show that ARRNs embrace the challenge posed by diverse resolutions with computational adaptivity, robustness, and compatibility with mainstream layers.

2025-07-16

Transactions on Machine Learning Research (accepté)

Curiosity-Driven Exploration via Temporal Contrastive Learning

Faisal Mohamed

Catherine Ji

Benjamin Eysenbach

Effective exploration in reinforcement learning requires keeping track not just of where the agent has been, but also of how the agent think… (voir plus)s about and represents the world: an agent should explore states that enable it to learn powerful representations. Temporal representations can include the information required to solve any potential task while avoiding the computational cost of reconstruction. In this paper, we propose an exploration method that uses temporal contrastive representations to drive exploration, maximizing coverage as seen through the lens of these temporal representations. We demonstrate complex exploration behaviors in locomotion, manipulation, and embodied-AI tasks, revealing previously unknown capabilities and behaviors once achievable only via extrinsic rewards.

2025-06-30

rl-conference.cc/RLC/2025/Workshop/RLBrew (publié)

Is Exploration or Optimization the Problem for Deep Reinforcement Learning?

2025-06-30

rl-conference.cc/RLC/2025/Workshop/Finding_the_Frame (publié)

Self-Predictive Representations for Combinatorial Generalization in Behavioral Cloning

Adriana Hugessen

Behavioral cloning (BC) methods trained with supervised learning (SL) are an effective way to learn policies from human demonstrations in do… (voir plus)mains like robotics. Goal-conditioning these policies enables a single generalist policy to capture diverse behaviors contained within an offline dataset. While goal-conditioned behavior cloning (GCBC) methods can perform well on in-distribution training tasks, they do not necessarily generalize zero-shot to tasks that require conditioning on novel state-goal pairs, i.e. combinatorial generalization. In part, this limitation can be attributed to a lack of temporal consistency in the state representation learned by BC; if temporally related states are encoded to similar latent representations, then the out-of-distribution gap for novel state-goal pairs would be reduced. Hence, encouraging this temporal consistency in the representation space should facilitate combinatorial generalization. Successor representations, which encode the distribution of future states visited from the current state, nicely encapsulate this property. However, previous methods for learning successor representations have relied on contrastive samples, temporal-difference (TD) learning, or both. In this work, we propose a simple yet effective representation learning objective,

2025-06-30

rl-conference.cc/RLC/2025/Workshop/RLBrew (publié)

What Matters for Maximizing Data Reuse In Value-based Deep Reinforcement Learning

Roger Creus Castanyer

Pablo Samuel Castro

A key ingredient for successfully applying deep reinforcement learning to challenging tasks is the effective use of data at scale. Although … (voir plus)originally deep RL algorithms achieved this by storing past experiences collected from a synchronous actor in an external replay memory [DQN; Mnih et al., 2013], follow-up works scaled training by collecting data asynchronously through distributed actors [R2D2; Kapturowski et al., 2018], and more recently by GPU-optimized parallelization [PQN; Gallici et al., 2024]. We argue that DQN, PQN, and R2D2 constitute a group of value-based methods for parallel training and study them to shed light on the dynamics induced by varying data collection schemes. We conduct a thorough empirical study to better understand these dynamics, and propose the Data Replay Ratio as a novel metric for quantifying data reuse. Our findings suggest that maximizing data reuse involves directly addressing the deadly triad: Q-lambda rollouts for reducing the bias from bootstrapping, the use of LayerNorm for stabilizing function approximation, and parallelized data collection for mitigating off-policy divergence.

2025-06-30

rl-conference.cc/RLC/2025/Workshop/Finding_the_Frame (publié)

Zero-Shot Constraint Satisfaction with Forward- Backward Representations

Adriana Hugessen

Harley Wiltzer

Cyrus Neary

Amy Zhang

Traditionally, constrained policy optimization with Reinforcement Learning (RL) requires learning a new policy from scratch for any new envi… (voir plus)ronment, goal or cost function, with limited generalization to new tasks and constraints. Given the sample inefficiency of many common deep RL methods, this procedure can be impractical for many real-world scenarios, particularly when constraints or tasks are changing. As an alternative, in the unconstrained setting, various works have sought to pre-train representations from offline datasets to accelerate policy optimization upon specification of a reward. Such methods can permit faster adaptation to new tasks in a given environment, dramatically improving sample efficiency. Recently, zero-shot policy optimization has been explored by leveraging a particular

2025-06-30

rl-conference.cc/RLC/2025/Workshop/RLBrew (publié)