The ability to process long contexts is crucial for many natural language processing tasks, yet it remains a significant challenge. While substantial progress has been made in enhancing the efficiency of attention mechanisms, there is still a gap in understanding how attention heads function in long-context settings. In this paper, we observe that while certain heads consistently attend to local information only, others swing between attending to local and long-context information depending on the query. This raises the question: can we identify which heads require long-context information to predict the next token accurately? We demonstrate that it's possible to predict which heads are crucial for long-context processing using only local keys. The core idea here is to exploit a simple model for the long-context scores via second moment approximations. These findings unveil simple properties of attention in the context of long sequences, and open the door to potentially significant gains in efficiency.

Gradient descent is the method of choice for training large artificial intelligence systems. As these systems become larger, a better understanding of the mechanisms behind gradient training would allow us to alleviate compute costs and help steer these systems away from harmful behaviors. To that end, we suggest utilizing the circuit perspective brought forward by mechanistic interpretability. After laying out our intuition, we illustrate how it enables us to design a curriculum for efficient learning in a controlled setting. The code is available at \url{https://github.com/facebookresearch/pal}.

Statistical learning in high-dimensional spaces is challenging without a strong underlying data structure. Recent advances with foundational models suggest that text and image data contain such hidden structures, which help mitigate the curse of dimensionality. Inspired by results from nonparametric statistics, we hypothesize that this phenomenon can be partially explained in terms of decomposition of complex tasks into simpler subtasks. In this paper, we present a controlled experimental framework to test whether neural networks can indeed exploit such ``hidden factorial structures.'' We find that they do leverage these latent patterns to learn discrete distributions more efficiently, and derive scaling laws linking model sizes, hidden factorizations, and accuracy. We also study the interplay between our structural assumptions and the models' capacity for generalization.

This paper introduces a visual sandbox designed to explore the training dynamics of a small-scale transformer model, with the embedding dimension constrained to $d=2$. This restriction allows for a comprehensive two-dimensional visualization of each layer's dynamics. Through this approach, we gain insights into training dynamics, circuit transferability, and the causes of loss spikes, including those induced by the high curvature of normalization layers. We propose strategies to mitigate these spikes, demonstrating how good visualization facilitates the design of innovative ideas of practical interest. Additionally, we believe our sandbox could assist theoreticians in assessing essential training dynamics mechanisms and integrating them into future theories. The code is available at https://github.com/facebookresearch/pal.

In the rapidly evolving field of artificial intelligence, Large Language Models (LLMs) have emerged as a pivotal component [45]. Their ability to understand, generate, and manipulate human language has opened up new avenues towards advanced machine intelligence. Interestingly, despite being primarily trained on next-token prediction tasks, LLMs are able to produce much more sophisticated answers when asked to generate steps of reasoning [30, 58]. This phenomenon, often referred to as Chain-of-Thought (CoT) reasoning, and illustrated on Table 1, appears paradoxical: on the one hand, LLMs are not explicitly programmed to reason; on the other hand, they are capable of following logical chains of thoughts to produce relatively complex answers. Table 1: Chain-of-Thought consists in eliciting reasoning steps before answering (A) a question (Q).

This work focuses on the training dynamics of one associative memory module storing outer products of token embeddings. We reduce this problem to the study of a system of particles, which interact according to properties of the data distribution and correlations between embeddings. Through theory and experiments, we provide several insights. In overparameterized regimes, we obtain logarithmic growth of the ``classification margins.'' Yet, we show that imbalance in token frequencies and memory interferences due to correlated embeddings lead to oscillatory transitory regimes. The oscillations are more pronounced with large step sizes, which can create benign loss spikes, although these learning rates speed up the dynamics and accelerate the asymptotic convergence. In underparameterized regimes, we illustrate how the cross-entropy loss can lead to suboptimal memorization schemes. Finally, we assess the validity of our findings on small Transformer models.

The combination of lightly supervised pre-training and online fine-tuning has played a key role in recent AI developments. These new learning pipelines call for new theoretical frameworks. In this paper, we formalize core aspects of weakly supervised and active learning with a simple problem: the estimation of the mode of a distribution using partial feedback. We show how entropy coding allows for optimal information acquisition from partial feedback, develop coarse sufficient statistics for mode identification, and adapt bandit algorithms to our new setting. Finally, we combine those contributions into a statistically and computationally efficient solution to our problem.

The number of sampling methods could be daunting for a practitioner looking to cast powerful machine learning methods to their specific problem. This paper takes a theoretical stance to review and organize many sampling approaches in the ``generative modeling'' setting, where one wants to generate new data that are similar to some training examples. By revealing links between existing methods, it might prove useful to overcome some of the current challenges in sampling with diffusion models, such as long inference time due to diffusion simulation, or the lack of diversity in generated samples.

The theory of statistical learning has focused on variational objectives expressed on functions. In this note, we discuss motivations to write similar objectives on measures, in particular to discuss out-of-distribution generalization and weakly-supervised learning. It raises a natural question: can one cast usual statistical learning results to objectives expressed on measures? Does the resulting construction lead to new algorithms of practical interest?

Large language models based on transformers have achieved great empirical successes. However, as they are deployed more widely, there is a growing need to better understand their internal mechanisms in order to make them more reliable. These models appear to store vast amounts of knowledge from their training data, and to adapt quickly to new information provided in their context or prompt. We study how transformers balance these two types of knowledge by considering a synthetic setup where tokens are generated from either global or context-specific bigram distributions. By a careful empirical analysis of the training process on a simplified two-layer transformer, we illustrate the fast learning of global bigrams and the slower development of an "induction head" mechanism for the in-context bigrams. We highlight the role of weight matrices as associative memories, provide theoretical insights on how gradients enable their learning during training, and study the role of data-distributional properties.