Paper: withholding training data only delays capabilities as general progress substitutes
Mark Marosi · hf · 2026-11-20
A preprint by Mark Marosi and Peter Antal (BME, Nov 19, 2026) tests whether withholding training data can actually prevent language models from acquiring skills.
- Core finding: a model's general progress on the rest of its training data substitutes for withheld data — filtering only delays a capability, it doesn't prevent it.
- Experiments withheld the training signal for individual skills in chess, Othello, molecular strings, bracket languages, and web text. The delay of a withheld skill in larger models was predictable from smaller ones.
- Key results: a chess model never explicitly trained on check responses ranked legal responses almost as well as a trained model; when later trained on the withheld data, it needed only a small fraction of the data required from initialization, and that fraction shrank as models improved. A control rule of equal difficulty was learned much more slowly; a withheld programming language in a 7B model was regained the same way.
- Implication: data filtering delays capabilities rather than blocking them, and the data needed to regain them decreases as models improve — an empirical challenge to filtering as a safety tool.
More from Safety
- Windows 11 Narrator 0day lets attackers bypass lockscreen and BitLocker in 5 minutes — rickasaurus · 2026-11-20
- Herbie Bradley: preventing model weight theft via SL5 security imposes huge costs — herbiebradley · 2026-11-20
- Polymarket puts DOJ criminally charging Nvidia this year at just 5% — Polymarket · 2026-11-20
- California launches investigation into Nvidia over allegedly misleading investors — Polymarket · 2026-11-20
- Rep. Ro Khanna urges California special session to pass AI safety legislation — trevposts · 2026-11-20
- Claude Code 2.1.329 removes built-in confirmation rule for high-risk actions — ClaudeCodeLog · 2026-11-20