Paper · Rough draft v0.1 · measured result · not peer reviewed · PDF, 10 pages
A third of forgetting is the arbitrary choice in the example
Rough draft of the paper. An exact solver makes every demonstration provably optimal, so only the arbitrary tie-break varies; that tie-break is a third of what the next task destroys.
Clark Whitehead · November 2026
Most tasks have several equally correct answers, and a demonstration has to show one of them. In grid worlds with an exact solver every demonstration is provably optimal, so the content of each demonstration can be held fixed while only that arbitrary pick, the convention, varies. Teaching a second task with demonstrations that break ties the opposite way from the first task's costs 0.10 of retention in 24 of 24 cells; under a clashing convention a third of all forgetting is the convention. A skill learned from all optimal actions at once is immune to the clash but no more durable when the conventions agree; choosing the demonstration the learner already prefers recovers exactly the convention effect; at matched KL divergence a convention-free skill loses half as much. A synthetic classifier reproduces the signature. 1,088 runs, pre-registered predictions, about $4 of GPU time.
Rough draft v0.1, November 2026 — measured result, not peer reviewed. Open before submission: a language-model replication and more than two conventions. The plain-language version is the research page.
More papers