Guillaume Lajoie

Biographie

Guillaume Lajoie est professeur agrégé au Département de mathématiques et de statistiques (DMS) de l'Université de Montréal et membre académique principal de Mila – Institut québécois d’intelligence artificielle. Il est titulaire d'une chaire CIFAR (CCAI Canada) ainsi que d'une chaire de recherche du Canada (CRC) en calcul et interfaçage neuronaux.

Ses recherches sont positionnées à l'intersection de l'IA et des neurosciences où il développe des outils pour mieux comprendre les mécanismes d'intelligence communs aux systèmes biologiques et artificiels. Les contributions de son groupe de recherche vont des progrès des paradigmes d'apprentissage à plusieurs échelles pour les grands systèmes artificiels aux applications en neurotechnologie. Dr. Lajoie participe activement aux efforts de développement responsables de l'IA, cherchant à identifier les lignes directrices et les meilleures pratiques pour l'utilisation de l'IA dans la recherche et au-delà.

Étudiants actuels

Federico Arangath Joseph

Collaborateur·rice de recherche - ETH Zurich

Rohan Banerjee

Collaborateur·rice alumni - Polytechnique

Visiteur de recherche indépendant

Superviseur⋅e principal⋅e :

Yoshua Bengio

Sangnie Bhardwaj

Doctorat - UdeM

Co-superviseur⋅e :

Hugo Larochelle

Colin Bredenberg

Postdoctorat - UdeM

Co-superviseur⋅e :

Leo Choiniere

Doctorat - UdeM

Olivier Codol

Postdoctorat - UdeM

Co-superviseur⋅e :

Doctorat - UdeM

Superviseur⋅e principal⋅e :

Leo Gagnon

Doctorat - UdeM

Tom George

Postdoctorat - McGill

Superviseur⋅e principal⋅e :

Juan Guerra

Maîtrise recherche - Polytechnique

Superviseur⋅e principal⋅e :

Nanda Harishankar Krishna

Doctorat - UdeM

Anna Jahn

Visiteur de recherche indépendant - McGill

Chen Jiang

Doctorat - McGill

Superviseur⋅e principal⋅e :

Paul Masset

Thomas Jiralerspong

Doctorat - UdeM

Co-superviseur⋅e :

Maîtrise recherche - UdeM

Co-superviseur⋅e :

Doctorat - McGill

Superviseur⋅e principal⋅e :

Mathys Loiselle

Stagiaire de recherche - Concordia

Co-superviseur⋅e :

Matt Perich

Ximeng Mao

Doctorat - UdeM

Co-superviseur⋅e :

Abdel Mfougouon Njupoun

Doctorat - UdeM

Co-superviseur⋅e :

Doctorat - UdeM

Co-superviseur⋅e :

Collaborateur·rice de recherche - UdeM

Mohammad Pezeshki

Collaborateur·rice de recherche

Superviseur⋅e principal⋅e :

Irina Rish

Julia Price

Maîtrise recherche - UdeM

Mauricio Rivera

Maîtrise recherche - UdeM

Superviseur⋅e principal⋅e :

Marco Bonizzato

Avery Ryoo

Doctorat - UdeM

Superviseur⋅e principal⋅e :

Doctorat - UdeM

Co-superviseur⋅e :

Lune Bellec

Ryan Vogt

Postdoctorat - UdeM

Doctorat - UdeM

Jieyu Zhao

Visiteur de recherche indépendant - University of South California

Apprentissage automatique pour la segmentation des différentes activations des fibres nerveuses à partir des signaux neuronaux du cerveau vers le corps

Billets de blogue

Représentation graphique d'un nerf vague

21 mai 2025

par

Param Raval

Olivier Tessier-Larivière

Pascal Fortier-Poisson

Blake Richards

Guillaume Lajoie

Lire l'article

13 juin 2024

Que nous apprennent les distributions des coefficients synaptiques au sujet de l’apprentissage dans le cerveau ?

par

Roman Pogodin

Jonathan Cornford

Arna Ghosh

Gauthier Gidel

Guillaume Lajoie

Blake Richards

Lire l'article

Publications

Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning

Seijin Kobayashi

Yanick Schimpf

Maximilian Schlegel

Angelika Steger

Maciej Wolczyk

Johannes Von Oswald

Nino Scherrer

Kaitlin Maile

Guillaume Lajoie

Rif A. Saurous

James Manyika

Blaise Aguera y Arcas

Alexander Meulemans

João Sacramento

Large-scale autoregressive models pretrained on next-token prediction and finetuned with reinforcement learning (RL) have achieved unprecede… (voir plus)nted success on many problem domains. During RL, these models explore by generating new outputs, one token at a time. However, sampling actions token-by-token can result in highly inefficient learning, particularly when rewards are sparse. Here, we show that it is possible to overcome this problem by acting and exploring within the internal representations of an autoregressive model. Specifically, to discover temporally-abstract actions, we introduce a higher-order, non-causal sequence model whose outputs control the residual stream activations of a base autoregressive model. On grid world and MuJoCo-based tasks with hierarchical structure, we find that the higher-order model learns to compress long activation sequence chunks onto internal controllers. Critically, each controller executes a sequence of behaviorally meaningful actions that unfold over long timescales and are accompanied with a learned termination condition, such that composing multiple controllers over time leads to efficient exploration on novel tasks. We show that direct internal controller reinforcement, a process we term "internal RL", enables learning from sparse rewards in cases where standard RL finetuning fails. Our results demonstrate the benefits of latent action generation and reinforcement in autoregressive models, suggesting internal RL as a promising avenue for realizing hierarchical RL within foundation models.

2025-12-22

ArXiv (prépublication)

Training neural networks from scratch in a videogame leads to brittle brain encoding

Basile Pinsard

Recent brain-encoding studies using videogame tasks suggest that the training objective of an artificial neural network plays a central role… (voir plus) in how well the network’s representations align with brain activity. This study investigates the alignment of artificial neural network activations with brain activity elicited by a video game task using models trained from scratch in controlled settings. We specifically compared three model training objectives: reinforcement learning, imitation learning, and a vision task, while accounting for other potential factors which may impact performance such as training data and model architecture. We tested models on brain encoding, i.e. their ability to predict functional magnetic resonance imaging (fMRI) signals acquired while human subjects played different levels of the video game Super Mario Bros. When tested on new playthroughs from the game levels seen at training, the reinforcement learning objective had a small but significant advantage in brain encoding, followed by the imitation learning and vision models. We hypothesized that brain-aligned representations would emerge only in task-competent models, and that the specific brain regions well encoded by a model would depend on the nature of the task it was trained on. While brain encoding did improve during model training, even an untrained model with matching architecture approached the performance of the best models. Contrary to our hypotheses, no model layers or specific training objectives aligned preferentially with specific brain areas. Large performance gaps also persisted in fully trained models across game levels, both those seen during training and entirely novel ones. Overall, even though reinforcement learning presented a small advantage to train brain encoding models for videogame data, all tested brain encoding models exhibited brittle performance with limited generalization both within- and out-of-distribution. Overall, our results suggest that training small artificial models from scratch is not sufficiently reliable, and that incorporating pretrained models such as foundation vision–action models may ultimately be necessary to support robust inferences about brain representations.

2025-12-01

bioRxiv (prépublication)

Embedded Universal Predictive Intelligence: a coherent framework for multi-agent learning

Alexander Meulemans

Rajai Nasser

Maciej Wolczyk

Marissa A. Weis

Seijin Kobayashi

Guillaume Lajoie

Angelika Steger

Marcus Hutter

James Manyika

Rif A. Saurous

João Sacramento

Blaise Aguera y Arcas

2025-11-26

ArXiv (prépublication)

Iterative Amortized Inference: Unifying In-Context Learning and Learned Optimizers

2025-10-13

ArXiv (prépublication)

Iterative Amortized Inference: Unifying In-Context Learning and Learned Optimizers

Modern learning systems increasingly rely on amortized learning - the idea of reusing computation or inductive biases shared across tasks to… (voir plus) enable rapid generalization to novel problems. This principle spans a range of approaches, including meta-learning, in-context learning, prompt tuning, learned optimizers and more. While motivated by similar goals, these approaches differ in how they encode and leverage task-specific information, often provided as in-context examples. In this work, we propose a unified framework which describes how such methods differ primarily in the aspects of learning they amortize - such as initializations, learned updates, or predictive mappings - and how they incorporate task data at inference. We introduce a taxonomy that categorizes amortized models into parametric, implicit, and explicit regimes, based on whether task adaptation is externalized, internalized, or jointly modeled. Building on this view, we identify a key limitation in current approaches: most methods struggle to scale to large datasets because their capacity to process task data at inference (e.g., context length) is often limited. To address this, we propose iterative amortized inference, a class of models that refine solutions step-by-step over mini-batches, drawing inspiration from stochastic optimization. Our formulation bridges optimization-based meta-learning with forward-pass amortization in models like LLMs, offering a scalable and extensible foundation for general-purpose task adaptation.

2025-10-13

ArXiv (prépublication)

Does learning the right latent variables necessarily improve in-context learning?

Large autoregressive models like Transformers can solve tasks through in-context learning (ICL) without learning new weights, suggesting ave… (voir plus)nues for efficiently solving new tasks. For many tasks, e.g., linear regression, the data factorizes: examples are independent given a task latent that generates the data, e.g., linear coefficients. While an optimal predictor leverages this factorization by inferring task latents, it is unclear if Transformers implicitly do so or if they instead exploit heuristics and statistical shortcuts enabled by attention layers. Both scenarios have inspired active ongoing work. In this paper, we systematically investigate the effect of explicitly inferring task latents. We minimally modify the Transformer architecture with a bottleneck designed to prevent shortcuts in favor of more structured solutions, and then compare performance against standard Transformers across various ICL tasks. Contrary to intuition and some recent works, we find little discernible difference between the two; biasing towards task-relevant latent variables does not lead to better out-of-distribution performance, in general. Curiously, we find that while the bottleneck effectively learns to extract latent task variables from context, downstream processing struggles to utilize them for robust prediction. Our study highlights the intrinsic limitations of Transformers in achieving structured ICL solutions that generalize, and shows that while inferring the right latents aids interpretability, it is not sufficient to alleviate this problem.

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

openreview.net

Does learning the right latent variables necessarily improve in-context learning?

Large autoregressive models like Transformers can solve tasks through in-context learning (ICL) without learning new weights, suggesting ave… (voir plus)nues for efficiently solving new tasks. For many tasks, e.g., linear regression, the data factorizes: examples are independent given a task latent that generates the data, e.g., linear coefficients. While an optimal predictor leverages this factorization by inferring task latents, it is unclear if Transformers implicitly do so or instead exploit heuristics and statistical shortcuts through attention layers. In this paper, we systematically investigate the effect of explicitly inferring task latents by minimally modifying the Transformer architecture with a bottleneck to prevent shortcuts and incentivize structured solutions. We compare it against standard Transformers across various ICL tasks and find that contrary to intuition and recent works, there is little discernible difference between the two; biasing towards task-relevant latent variables does not lead to better out-of-distribution performance, in general. Curiously, we find that while the bottleneck effectively learns to extract latent task variables from context, downstream processing struggles to utilize them for robust prediction. Our study highlights the intrinsic limitations of Transformers in achieving structured ICL solutions that generalize, and shows that while inferring the right latents aids interpretability, it is not sufficient to alleviate this problem.

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

In-context learning and Occam's razor

A central goal of machine learning is generalization. While the No Free Lunch Theorem states that we cannot obtain theoretical guarantees fo… (voir plus)r generalization without further assumptions, in practice we observe that simple models which explain the training data generalize best: a principle called Occam's razor. Despite the need for simple models, most current approaches in machine learning only minimize the training error, and at best indirectly promote simplicity through regularization or architecture design. Here, we draw a connection between Occam's razor and in-context learning: an emergent ability of certain sequence models like Transformers to learn at inference time from past observations in a sequence. In particular, we show that the next-token prediction loss used to train in-context learners is directly equivalent to a data compression technique called prequential coding, and that minimizing this loss amounts to jointly minimizing both the training error and the complexity of the model that was implicitly learned from context. Our theory and the empirical experiments we use to support it not only provide a normative account of in-context learning, but also elucidate the shortcomings of current in-context learning methods, suggesting ways in which they can be improved. We make our code available at https://github.com/3rdCore/PrequentialCode.

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

openreview.net

In-Context Learning and Occam’s Razor

A central goal of machine learning is generalization. While the No Free Lunch Theorem states that we cannot obtain theoretical guarantees fo… (voir plus)r generalization without further assumptions, in practice we observe that simple models which explain the training data generalize best—a principle called Occam’s razor. Despite the need for simple models, most current approaches in machine learning only minimize the training error, and at best indirectly promote simplicity through regularization or architecture design. Here, we draw a connection between Occam’s razor and in-context learning—an emergent ability of certain sequence models like Transformers to learn at inference time from past observations in a sequence. In particular, we show that the next-token prediction loss used to train in-context learners is directly equivalent to a data compression technique called prequential coding, and that minimizing this loss amounts to jointly minimizing both the training error and the complexity of the model that was implicitly learned from context. Our theory and the empirical experiments we use to support it not only provide a normative account of in-context learning, but also elucidate the shortcomings of current in-context learning methods, suggesting ways in which they can be improved. We make our code available at https://github.com/3rdCore/PrequentialCode.

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

Towards a Formal Theory of Representational Compositionality

Compositionality is believed to be fundamental to intelligence. In humans, it underlies the structure of thought and language. In AI, it ena… (voir plus)bles a powerful form of out-of-distribution generalization, in which a model systematically adapts to novel combinations of known concepts. However, while we have strong intuitions about what compositionality is, we lack satisfying formal definitions for it. Here, we propose such a definition called representational compositionality that is conceptually simple, quantitative, and grounded in algorithmic information theory. Intuitively, representational compositionality states that a compositional representation is both expressive and describable as a simple function of parts. We validate our definition on both real and synthetic data, and show how it unifies disparate intuitions from across the literature in both AI and cognitive science. We hope that our definition can inspire the design of novel, theoretically-driven models that better capture the mechanisms of compositional thought. We make our code available at https://github.com/EricElmoznino/complexity_compositionality.

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

Towards a Formal Theory of Representational Compositionality

2025-10-06

Proceedings of the 42nd International Conference on Machine Learning (publié)

openreview.net

Recursive Self-Aggregation Unlocks Deep Thinking in Large Language Models

Johan Samir Obando Ceron

Yoshua Bengio

Brian R. Bartoldson

Bhavya Kailkhura

Guillaume Lajoie

Glen Berseth

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.

2025-09-30

ArXiv (prépublication)