Memorization: A Close Look at Books
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866915585763311616 |
|---|---|
| author | Ma, Iris Domingo, Ian Krone-Martins, Alberto Baldi, Pierre Lopes, Cristina V. |
| author_facet | Ma, Iris Domingo, Ian Krone-Martins, Alberto Baldi, Pierre Lopes, Cristina V. |
| contents | To what extent can entire books be extracted from LLMs? Using the Llama 3 70B family of models, and the "prefix-prompting" extraction technique, we were able to auto-regressively reconstruct, with a very high level of similarity, one entire book (Alice's Adventures in Wonderland) from just the first 500 tokens. We were also able to obtain high extraction rates on several other books, piece-wise. However, these successes do not extend uniformly to all books. We show that extraction rates of books correlate with book popularity and thus, likely duplication in the training data.
We also confirm the undoing of mitigations in the instruction-tuned Llama 3.1, following recent work (Nasr et al., 2025). We further find that this undoing comes from changes to only a tiny fraction of weights concentrated primarily in the lower transformer blocks. Our results provide evidence of the limits of current regurgitation mitigation strategies and introduce a framework for studying how fine-tuning affects the retrieval of verbatim memorization in aligned LLMs. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_12549 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Memorization: A Close Look at Books Ma, Iris Domingo, Ian Krone-Martins, Alberto Baldi, Pierre Lopes, Cristina V. Computation and Language Artificial Intelligence Machine Learning To what extent can entire books be extracted from LLMs? Using the Llama 3 70B family of models, and the "prefix-prompting" extraction technique, we were able to auto-regressively reconstruct, with a very high level of similarity, one entire book (Alice's Adventures in Wonderland) from just the first 500 tokens. We were also able to obtain high extraction rates on several other books, piece-wise. However, these successes do not extend uniformly to all books. We show that extraction rates of books correlate with book popularity and thus, likely duplication in the training data. We also confirm the undoing of mitigations in the instruction-tuned Llama 3.1, following recent work (Nasr et al., 2025). We further find that this undoing comes from changes to only a tiny fraction of weights concentrated primarily in the lower transformer blocks. Our results provide evidence of the limits of current regurgitation mitigation strategies and introduce a framework for studying how fine-tuning affects the retrieval of verbatim memorization in aligned LLMs. |
| title | Memorization: A Close Look at Books |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2504.12549 |