REBEL: Hidden Knowledge R ecovery via E volutionary- B ased E valuation L oop
Patryk Rybak, Paweł Batorski, Paul Swoboda, Przemysław Spurek · 2026
Read REBEL: Hidden Knowledge R ecovery via E volutionary- B ased E valuation L oop by Patryk Rybak, Paweł Batorski, Paul Swoboda, Przemysław Spurek (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-2602-06248.txt.
Description
Deleting sensitive or copyrighted material from a trained language model is called machine unlearning, but checking that the knowledge is truly gone is hard. Patryk Rybak and 3 co-authors present REBEL, a system that hunts for leftovers. Standard checks use tame questions, which can mistake surface silence for real forgetting.
REBEL evolves adversarial prompts, breeding new attacks from ones that partly worked. On two standard benchmarks, it pulled supposedly forgotten knowledge back out of models, with success rates far above fixed test prompts. The results span many unlearning algorithms. The authors warn that current methods may offer only a thin layer of protection.
Catalog record
| CREATOR | Patryk Rybak, Paweł Batorski, Paul Swoboda, Przemysław Spurek |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06248