The Geometry of Sequential Learning: Lie-Bracket Prediction of Transfer Order
John Sweeney · 2026
Read The Geometry of Sequential Learning: Lie-Bracket Prediction of Transfer Order by John Sweeney (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-24993.txt.
Description
John Sweeney shows that the order in which training data is presented to a model can be predicted from geometry. Sequential learning is order dependent: the number of possible curricula explodes. Sweeney proves the order effect is governed by a computable quantity, the Lie bracket commutator of gradient update fields. It yields a pairwise score for whether source A should precede B or the reverse. From that primitive he builds a tournament method that ranks all sources cheaply.
The tournament needs one Hessian vector product per source and a fast sort, never the full pairwise matrix. Empirically the planner shines: about 98 percent pairwise accuracy for single step instruction tuning and preference optimisation, and about 73 percent twenty steps out. It recovers the best of all six schedules for three sources in about 88 percent of trials. On 85 programming language domains aimed at a Python target it ranks in the 99th sampled percentile.
Catalog record
| CREATOR | John Sweeney |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 79 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.24993