On the "Induction Bias" in Sequence Models

TL;DR

This paper investigates the data efficiency and state-tracking capabilities of transformer-based models versus RNNs, revealing transformers require more data and learn length-specific solutions, highlighting ongoing challenges in sequence modeling.

Contribution

The study provides a large-scale empirical comparison showing transformers' limitations in data efficiency and weight sharing for state tracking compared to RNNs.

Findings

Transformers need more data as sequence length increases.

Transformers learn length-specific solutions with little weight sharing.

RNNs effectively share weights across sequence lengths.

Abstract

Despite the remarkable practical success of transformer-based language models, recent work has raised concerns about their ability to perform state tracking. In particular, a growing body of literature has shown this limitation primarily through failures in out-of-distribution (OOD) generalization, such as length extrapolation. In this work, we shift attention to the in-distribution implications of these limitations. We conduct a large-scale experimental study of the data efficiency of transformers and recurrent neural networks (RNNs) across multiple supervision regimes. We find that the amount of training data required by transformers grows much more rapidly with state-space size and sequence length than for RNNs. Furthermore, we analyze the extent to which learned state-tracking mechanisms are shared across different sequence lengths. We show that transformers exhibit negligible or even detrimental weight sharing across lengths, indicating that they learn length-specific solutions in isolation. In contrast, recurrent models exhibit effective amortized learning by sharing weights across lengths, allowing data from one sequence length to improve performance on others. Together, these results demonstrate that state tracking remains a fundamental challenge for transformers, even when training and evaluation distributions match.