Razvan Pascanu

Blaise Agüera y Arcas

João Sacramento

Sequence modeling is currently dominated by causal transformer architectures that use softmax self-attention. Although widely adopted, trans… (voir plus)formers require scaling memory and compute linearly during inference. A recent stream of work linearized the softmax operation, resulting in powerful recurrent neural network (RNN) models with constant memory and compute costs such as DeltaNet, Mamba or xLSTM. These models can be unified by noting that their recurrent layer dynamics can all be derived from an in-context regression objective, approximately optimized through an online learning rule. Here, we join this line of work and introduce a numerically stable, chunkwise parallelizable version of the recently proposed Mesa layer (von Oswald et al., 2024), and study it in language modeling at the billion-parameter scale. This layer again stems from an in-context loss, but which is now minimized to optimality at every time point using a fast conjugate gradient solver. Through an extensive suite of experiments, we show that optimal test-time training enables reaching lower language modeling perplexity and higher downstream benchmark performance than previous RNNs, especially on tasks requiring long context understanding. This performance gain comes at the cost of additional flops spent during inference time. Our results are therefore intriguingly related to recent trends of increasing test-time compute to improve performance -- here by spending compute to solve sequential optimization problems within the neural network itself.

2025-06-04

ArXiv (prépublication)

Unpacking Softmax: How Temperature Drives Representation Collapse, Compression, and Generalization

Wojciech Masarczyk

Mateusz Ostaszewski

Tin Sum Cheng

Tomasz Trzci'nski

Aurélien Lucchi

The softmax function is a fundamental building block of deep neural networks, commonly used to define output distributions in classification… (voir plus) tasks or attention weights in transformer architectures. Despite its widespread use and proven effectiveness, its influence on learning dynamics and learned representations remains poorly understood, limiting our ability to optimize model behavior. In this paper, we study the pivotal role of the softmax function in shaping the model's representation. We introduce the concept of rank deficit bias - a phenomenon in which softmax-based deep networks find solutions of rank much lower than the number of classes. This bias depends on the softmax function's logits norm, which is implicitly influenced by hyperparameters or directly modified by softmax temperature. Furthermore, we demonstrate how to exploit the softmax dynamics to learn compressed representations or to enhance their performance on out-of-distribution data. We validate our findings across diverse architectures and real-world datasets, highlighting the broad applicability of temperature tuning in improving model performance. Our work provides new insights into the mechanisms of softmax, enabling better control over representation learning in deep neural networks.

2025-06-01

ArXiv (prépublication)

On the generalization of language models from in-context learning and finetuning: a controlled study

Andrew Lampinen

Arslan Chaudhry

Stephanie C.Y. Chan

Cody Wild

Diane Wan

Alexander Y. Ku

Alex Ku

Jörg Bornschein

Murray P. Shanahan

James L McClelland

2025-04-30

ArXiv (prépublication)

LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilities

Thomas Schmied

Jörg Bornschein

Jordi Grau-Moya

Markus Wulfmeier

2025-04-21

ArXiv (prépublication)

Why do LLMs attend to the first token?

Federico Barbero

'Alvaro Arroyo

Xiangming Gu

Christos Perivolaropoulos

Michael M. Bronstein

Petar Veličković

2025-04-02

ArXiv (prépublication)

NoProp: Training Neural Networks without Back-propagation or Forward-propagation

Qinyu Li

Yee Whye Teh

2025-03-30

ArXiv (prépublication)

How do language models learn facts? Dynamics, curricula and hallucinations

Nicolas Zucchet

Jörg Bornschein

Stephanie Chan

Andrew Lampinen

Soham De

2025-03-26

ArXiv (prépublication)

Disentangling the Causes of Plasticity Loss in Neural Networks

Clare Lyle

Zeyu Zheng

Khimya Khetarpal

Hado van Hasselt

James Martens

Will Dabney

2025-02-16

Proceedings of The 3rd Conference on Lifelong Learning Agents (publié)

proceedings.mlr.press

From Markov to Laplace: How Mamba In-Context Learns Markov Chains

Marco Bondaschi

Nived Rajaraman

Xiuying Wei

Kannan Ramchandran

Caglar Gulçehre

Michael C. Gastpar

Ashok Vardhan Makkuva

2025-02-13

ArXiv (prépublication)

Maxwell's Demon at Work: Efficient Pruning by Leveraging Saturation of Neurons

When training neural networks, dying neurons -- units becoming inactive or saturated -- are traditionally seen as harmful. This paper sheds … (voir plus)new light on this phenomenon. By exploring the impact of various hyperparameter configurations on dying neurons during training, we gather insights on how to improve upon sparse training approaches to pruning. We introduce Demon Pruning (DemP), a method that controls the proliferation of dead neurons through a combination of noise injection on active units and a one-cycle schedule regularization strategy, dynamically leading to network sparsity. Experiments on CIFAR-10 and ImageNet datasets demonstrate that DemP outperforms existing dense-to-sparse structured pruning methods, achieving better accuracy-sparsity tradeoffs and accelerating training by up to 3.56

2025-02-12

TMLR (accepté)

openreview.net

Agency Is Frame-Dependent

David Abel

Andre Barreto

Michael Bowling

Will Dabney

Shi Dong

Steven Stenberg Hansen

Anna Harutyunyan

Khimya Khetarpal

Clare Lyle

Georgios Piliouras

Doina Precup

Jonathan Richens

Mark Rowland

Tom Schaul

Satinder Singh

Agency is a system's capacity to steer outcomes toward a goal, and is a central topic of study across biology, philosophy, cognitive science… (voir plus), and artificial intelligence. Determining if a system exhibits agency is a notoriously difficult question: Dennett (1989), for instance, highlights the puzzle of determining which principles can decide whether a rock, a thermostat, or a robot each possess agency. We here address this puzzle from the viewpoint of reinforcement learning by arguing that agency is fundamentally frame-dependent: Any measurement of a system's agency must be made relative to a reference frame. We support this claim by presenting a philosophical argument that each of the essential properties of agency proposed by Barandiaran et al. (2009) and Moreno (2018) are themselves frame-dependent. We conclude that any basic science of agency requires frame-dependence, and discuss the implications of this claim for reinforcement learning.

2025-02-05

ArXiv (prépublication)

Attention as a Hypernetwork

Simon Schug

Seijin Kobayashi

Yassir Akram

João Sacramento

Transformers can under some circumstances generalize to novel problem instances whose constituent parts might have been encountered during t… (voir plus)raining, but whose compositions have not. What mechanisms underlie this ability for compositional generalization? By reformulating multi-head attention as a hypernetwork, we reveal that a composable, low-dimensional latent code specifies key-query specific operations. We find empirically that this latent code is predictive of the subtasks the network performs on unseen task compositions, revealing that latent codes acquired during training are reused to solve unseen problem instances. To further examine the hypothesis that the intrinsic hypernetwork of multi-head attention supports compositional generalization, we ablate whether making the hypernetwork-generated linear value network nonlinear strengthens compositionality. We find that this modification improves compositional generalization on abstract reasoning tasks. In particular, we introduce a symbolic version of the Raven's Progressive Matrices human intelligence test, which gives us precise control over the problem compositions encountered during training and evaluation. We demonstrate on this task how scaling model size and data enables compositional generalization in transformers and gives rise to a functionally structured latent space.

2024-12-31

ICLR (publié)