This paper discusses the internal behavior of Transformer language models.Many recent pre-trained models have been reported to exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle Transformer layers, despite a disproportionately large “jump” in the angular distance occurring in or around the final Transformer layer.To characterize this, we first introduce a quantitative metric for the jump strength around the final layer, and then demonstrate its prevalence across many open-weight models, as well as its amplification throughout pre-training.Assuming such jumps indicate an undesirable property, we propose the jump-suppressing regularizer (JREG) which penalizes this jump during pre-training, thereby encouraging more balanced capability usage across the middle layers.Empirical evaluations of three model sizes of Llama-based models, trained with the proposed JREG method, reveal improved task performance compared to the baseline without altering the model architecture.

Many linguistic expressions have idiomatic and literal interpretations, and the automatic distinction of these two interpretations has been studied for decades. Recent research has shown that contextualized word embeddings derived from masked language models (MLMs) can give promising results for idiom token classification. This indicates that contextualized word embedding alone contains information about whether the word is being used in a literal sense or not. However, we believe that more types of information can be derived from MLMs and that leveraging such information can improve idiom token classification. In this paper, we leverage three types of embeddings from MLMs; uncontextualized token embeddings and masked token embeddings in addition to the standard contextualized word embeddings and show that the newly added embeddings significantly improve idiom token classification for both English and Japanese datasets.