Generalization Dynamics of LM Pre-training ("Discovering Mode-Hopping")

Summary

The usual mental model -- that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training -- is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like computations ("mode-hopping"). Across a toy eval suite, LMs suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment -- then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and cannot be fixed by checkpoint averaging; the authors frame it as a capacity-allocation problem in which generalizable circuits compete with early-learned shallow circuits, with each pre-training window's data deciding which wins. Two applications: (1) selecting intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than final or mid-training checkpoints; (2) selecting pre-training data that controls and stabilizes generalization dynamics. Concrete numbers from the announcement: OLMo3-32B went from 81% accuracy to 0%, then back to 81.7% within 40B training tokens; an earlier 4.5T-token checkpoint beat a 4.9T one on GPQA transfer after math fine-tuning (36.3% vs 29.8%) and robustness to alignment attacks (53% vs 21%). Code/eval suite: GDsuite (GitHub).

Keywords

pretraining · generalization · mode-hopping · evaluation · checkpoints

Topics

pretraining, generalization, mode-hopping, evaluation

Research notes

• Discovery: alphaXiv (@askalphaxiv) 2026-09-30: https://x.com/askalphaxiv/status/2105350957609660836?s=20
• Tweet title: "Discovering Mode-Hopping: Large Language Models Repeatedly Hop between Pattern Matching and Generalization"; alphaXiv page title: "Generalization Dynamics of LM Pre-training" (2609.33150).
• Affiliations shown: UC Berkeley, Stanford, DeepMind (per-author mapping not given).
• Code: https://github.com/Jiaxin-Wen/GDsuite
• alphaXiv: https://www.alphaxiv.org/abs/2609.33150
• No license stated on the alphaXiv page.
• Connects to the collection's pretraining, generalization, and eval-suite entries.