Generalization Dynamics of LM Pretraining

Pretraining was not a smooth evolution from pattern imitation to generalization. Tracking intermediate checkpoints of the open models OLMo3 (7B, 32B) and Apertus (8B, 70B), the models repeatedly showed mode-hopping, latching onto in-context patterns only to suddenly snap back to generalizing circuits. Both models were trained up to 9–90x the Chinchilla-optimal budget, and this phenomenon could not be explained by simple undertraining.

The evaluation suite consisted of 6 toy evals plus 2 additional fine-tuning evaluations.

• Flipped / repetitive / successive answer ICL: checks whether the model follows incorrect labels or repetitive/sequential patterns, or actually solves the task.
• Truthy answer: checks whether the model chooses plausible-sounding sentences over factual accuracy.
• Intuitive answers: checks whether the model relies on System 1 or System 2.
• Multi-hop persona QA, out-of-context reasoning, emergent misalignment: checks whether the model connects scattered facts into coherent generalization.

The results were abrupt. For example, OLMo3 32B rose to 81% at 2.17T tokens on an evaluation where the answer increases by +1, then collapsed to 0% at 2.19T tokens, before recovering to 81.7% again at 2.21T tokens. Neither a single gradient step nor checkpoint averaging could adequately explain or fully fix this shift, and smaller models showed patterns of either slowly wobbling toward generalization or getting stuck in pattern imitation.

The key interpretation is capacity allocation. Within limited capacity, shallow circuits learned early compete with generalizing circuits, and the data in each pre-training window determines which side wins.

This perspective leads to three practical implications. Some intermediate checkpoints generalized better than the final checkpoint on GPQA and in subsequent alignment training, and pre-training data selection could be used to regulate and stabilize generalization dynamics. It also revealed that generalization predictors show opposite correlations depending on the layer, and that simple complexity metrics alone are insufficient to fully explain generalization.