Transformers Can Achieve Length Generalization But Not Robustly
This addresses the problem of length generalization in language models for researchers, showing it is achievable but not robust, which is incremental as it builds on prior work on Transformers and generalization challenges.
The paper investigates Transformers' ability to generalize to longer sequences than seen during training, using integer addition as a test case, and finds that with optimal data format and position encodings, they can extrapolate to lengths 2.5 times the training input, but this generalization is fragile and highly sensitive to factors like initialization and data order.
Length generalization, defined as the ability to extrapolate from shorter training sequences to longer test ones, is a significant challenge for language models. This issue persists even with large-scale Transformers handling relatively straightforward tasks. In this paper, we test the Transformer's ability of length generalization using the task of addition of two integers. We show that the success of length generalization is intricately linked to the data format and the type of position encoding. Using the right combination of data format and position encodings, we show for the first time that standard Transformers can extrapolate to a sequence length that is 2.5x the input length. Nevertheless, unlike in-distribution generalization, length generalization remains fragile, significantly influenced by factors like random weight initialization and training data order, leading to large variances across different random seeds.