Paper: LLMs repeatedly hop between shallow heuristics and generalization during pre-training
thomasahle · x · 2026-10-12
In "Generalization Dynamics of LM Pre-Training", Jiaxin Wen, Zhengxuan Wu, Dawn Song and Lijie Chen show that LLMs don't stably retain reasoning during pre-training: they repeatedly and suddenly hop between shallow pattern-matching heuristics and generalizable computation, even when pre-training loss and downstream capability curves look stable.
Per the paper's figure, the authors distinguish three generalization dynamics types (frequent oscillation, sustained oscillation, sparse spikes), comparing large and small models against Chinchilla-optimal points: stable surface metrics don't imply the internal computation has converged to a generalizing solution.
More from Research
- OpenAI paper claims proof that pi's irrationality exponent is exactly 2 — TinfoilTricorn · 2026-10-12
- Byteification retrofits token-based LLMs to work at byte level, fixing letter-counting failures — PontiEdoardo · 2026-10-12
- No orbits, just probability clouds: exact |psi|^2 isosurfaces of hydrogen wavefunctions — TinfoilTricorn · 2026-10-12
- Nobel economics prize goes to Robert E. Hall and Michael Woodford for microfoundations of macro stabilization — AnnaCiaunica · 2026-10-12
- Counterexample to Vaught's conjecture formalized and verified in Lean — roydanroy · 2026-10-12
- METR's AI time-horizon chart hasn't updated since May, leaving frontier model gains unmeasured — haider1 · 2026-10-12