Papers
arxiv:2611.24844

General progress substitutes for specific training data in language models

Published on Nov 18
· Submitted by
Mark Marosi
on Nov 19
Authors:

Abstract

Withholding training data delays but does not prevent language models from acquiring skills, and models can regain withheld capabilities with less data as they improve.

Language models acquire their capabilities from training data, and developers withhold data during training to keep models from learning capabilities that could be misused. How long this delays a capability, and how much data is needed to acquire it later, is unknown. Here we show that the general progress a model makes on the rest of its training data substitutes for the data that was withheld. We withheld the training signal for individual skills in models of chess, Othello, molecular strings, bracket languages and web text. The delay of a withheld skill in larger models was predicted from smaller ones. A chess model never explicitly trained to respond to check ranked the legal responses almost as well as a model that was. When models were then trained 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. A programming language withheld from the training text of a 7-billion-parameter model was regained in the same way. Data filtering therefore delays a capability, and the data required to regain it decreases as models improve.

Community

Paper author Paper submitter

Developers can limit the capabilities of AI models by removing certain topics from training data, but they’ve never known for how long this slows development. Now, we’ve found that general intelligence compensates for these exclusions.

We hid specific skills from AI models trained on chess, Othello, molecular strings, bracket languages, and web text. We could predict how long a specific skill would be delayed in a large model by observing small ones, but what surprised us most was how close these models came to developing skills without any data. A chess model with no data to respond to check still ranked legal moves almost as well as a trained model!
When we gave the models the withheld data, they acquired the withheld skill with a tiny fraction of the data needed by an untrained model. The more intelligent the model, the less data was needed! If we gave the models an equally difficult, unrelated rule, they were much slower to learn it. We found the same pattern at scale: a 7-billion-parameter model had a programming language withheld from its training data, but recovered it when exposed.

We conclude that filtering training data doesn’t exclude capabilities forever but instead delays them. As model intelligence rises, the data needed to develop these skills diminishes.

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Sign up or log in to comment

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2611.24844 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2611.24844 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2611.24844 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.