GENERALIZATION DYNAMICS OF LM PRE-TRAINING

Summary from the TLDR AI newsletter

During pre-training, models can abruptly switch response patterns, a phenomenon termed MODE-HOPPING. Tests reveal this behavior in OLMo3-32B during arithmetic tasks, showing sudden shifts between pattern-following and correct task inference as training progresses. Controlled experiments suggest competing shallow and generalizable circuits within models, revealing that longer training doesn't always guarantee better generalization.