paper web signal

Data filtering only delays AI skills, new arxiv paper finds

TL;DR

  • The paper tested five domains: chess, Othello, molecular strings, bracket languages and web text, finding withheld skills recover with less data as models grow.
  • A chess model never trained to respond to check ranked legal responses almost as well as a model that was, the authors report.
  • A programming language withheld from a 7-billion-parameter model was regained the same way; a control rule of equal difficulty was learned much more slowly.

A chess model that was never explicitly trained to respond to check still ranked the legal responses "almost as well as a model that was." That is the headline finding of a new arxiv paper by Mark Marosi and Peter Antal, which tests whether withholding training data prevents a model from acquiring a capability, or just delays it.

The authors withheld the training signal for individual skills across chess, Othello, molecular strings, bracket languages and web text, and separately withheld a programming language from a 7-billion-parameter model. The result in each case was the same. "Data filtering therefore delays a capability," the paper concludes, "and the data required to regain it decreases as models improve."

The delay was predictable. "The delay of a withheld skill in larger models was predicted from smaller ones," the authors write, meaning the window a filter buys shrinks on a curve a lab could in principle estimate before training a frontier model. When the authors then trained models on the withheld data, "they acquired the skill with a small fraction of the data needed from initialization, and this fraction decreased as the models progressed." A control rule of equal difficulty was learned much more slowly, the authors' argument that general training progress, not the specific data, is what carries the hidden skill back.

The abstract reports no per-skill numbers and names no safety-critical capabilities among the test domains; the test set is the games, the strings and the programming language.