IteRABRe: Ite rative R ecovery- A ided B lock Re duction
Haryo Akbarianto Wibowo, Haiyue Song, Hideki Tanaka, Masao Utiyama, Alham Fikri Aji, Raj Dabre · 2025
Read IteRABRe: Ite rative R ecovery- A ided B lock Re duction by Haryo Akbarianto Wibowo, Haiyue Song, Hideki Tanaka, Masao Utiyama, Alham Fikri Aji, Raj Dabre (2025) 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-2503-06291.txt.
Description
Haryo Akbarianto Wibowo and co-authors present IteRABRe, a lean way to shrink large language models. Their 2025 paper works on block pruning, cutting whole chunks from a network, a simple compression route that often damages skill or demands heavy retraining to repair.
Their method prunes in rounds and recovers skill after each round, using only two and a half million tokens for the repair steps. On two well-known open models it beat rival pruning methods by about three points on average. Language skills held up best of all, scoring about five points above the rivals, and skill across languages survived the cuts. The authors also chart how the two models differ under pruning.
Catalog record
| CREATOR | Haryo Akbarianto Wibowo, Haiyue Song, Hideki Tanaka, Masao Utiyama, Alham Fikri Aji, Raj Dabre |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-06291