General learned delegation by clones

Darren Li, Meiqi Chen, Chenze Shao, Fandong Meng, Jie Zhou · 2026

Read General learned delegation by clones by Darren Li, Meiqi Chen, Chenze Shao, Fandong Meng, Jie Zhou (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-2602-13262.txt.

Description

Giving a language model more thinking time helps, but running one long chain or many blind samples wastes compute under a fixed budget. SELFCEST teaches a base model to split itself into clones, copies with the same weights that work in parallel contexts and share what they learn. A learned controller divides both text output and context space across branches, trained end to end under one overall task reward.

Across hard math tests and long-context questions needing several hops, the clone approach beat single-model baselines at the same compute budget and held up on new task types. The idea reframes test-time compute as a delegation problem: one model that learns to manage its own team of copies. Darren Li leads the work with four co-authors.

Catalog record

CREATORDarren Li, Meiqi Chen, Chenze Shao, Fandong Meng, Jie Zhou
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-13262

Browse the full library · About the Tower of Babel