Contextual Counting: A Mechanistic Study of Transformers on a Quantitative Task

TL;DR

This paper introduces the contextual counting task to analyze Transformer behavior in quantitative tasks, revealing that causal attention and absence of positional embeddings improve accuracy and interpretability.

Contribution

It provides a theoretical and empirical study of Transformers on a new quantitative task, highlighting the impact of attention types and positional encodings.

Findings

Causal attention outperforms non-causal in the task.

No positional embeddings yield the best accuracy.

Rotary embeddings are competitive and easier to train.

Abstract

Transformers have revolutionized machine learning across diverse domains, yet understanding their behavior remains crucial, particularly in high-stakes applications. This paper introduces the contextual counting task, a novel toy problem aimed at enhancing our understanding of Transformers in quantitative and scientific contexts. This task requires precise localization and computation within datasets, akin to object detection or region-based scientific analysis. We present theoretical and empirical analysis using both causal and non-causal Transformer architectures, investigating the influence of various positional encodings on performance and interpretability. In particular, we find that causal attention is much better suited for the task, and that no positional embeddings lead to the best accuracy, though rotary embeddings are competitive and easier to train. We also show that out of distribution performance is tightly linked to which tokens it uses as a bias term.