GENERALIZATION DYNAMICS OF LM PRE-TRAINING

Summary from the TLDR AI newsletter

Language models (LMs) undergo unpredictable switches between parroting patterns and exhibiting adaptive intelligence during pre-training, a phenomenon termed "mode-hopping." This behavior cannot be corrected by standard optimization techniques and presents as a competition for model capacity, influenced by data from each training window. Researchers propose leveraging these dynamics to better select pre-training checkpoints, curate data for stable generalization, and evaluate metrics predicting LM behavior.