Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale

James A. Michaelov, Roger P. Levy, Ben Bergen

Keywords: language models, training dynamics, n-grams, semantic similarity, heuristics, shortcuts

TL;DR: Transformer, Mamba, and RWKV language models show consistent patterns of change in behavior over the course of training

Abstract: We show that across architecture (Transformer vs. Mamba vs. RWKV), training dataset (OpenWebText vs. The Pile), and scale (14 million parameters to 12 billion parameters), autoregressive language models exhibit highly consistent patterns of change in their behavior over the course of pretraining. Based on our analysis of over 1,400 language model checkpoints on over 110,000 tokens of English, we find that up to 98% of the variance in language model behavior at the word level can be explained by three simple heuristics: the unigram probability (frequency) of a given word, the n-gram probability of the word, and the semantic similarity between the word and its context. Furthermore, we see consistent behavioral phases in all language models, with their predicted probabilities for words overfitting to those words' n-gram probabilities for increasing over the course of training. Taken together, these results suggest that learning in neural language models may follow a similar trajectory irrespective of model details.

Scope Confirmation: To the best of my judgment, this submission falls within the scope of CoNLL.

Primary Area Selection: Theoretical Analysis and Interpretation of ML Models for NLP

Secondary Area Selection: Computational Psycholinguistics, Cognition and Linguistics, Language Acquisition, Learning, Emergence, and Evolution

Use Of Generative Artificial Intelligence Tools: No, not at all

Data Collection From Human Subjects: No

Submission Type: Non-archival (fill out the field below).

Publication Data: NeurIPS 2025: https://openreview.net/forum?id=HenpVfO3Wp

Submission Number: 139