Hard Constraints Meet Soft Generation: Guaranteed Feasibility for LLM-based Combinatorial Optimization

Yang Liu, Chuan Zhou, Yancheng Chen, Shuai Zhang, Xixun Lin, Xiaoqing Wang · 2026

Read Hard Constraints Meet Soft Generation: Guaranteed Feasibility for LLM-based Combinatorial Optimization by Yang Liu, Chuan Zhou, Yancheng Chen, Shuai Zhang, Xixun Lin, Xiaoqing Wang (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-2602-01090.txt.

Description

Yang Liu and co-authors build FALCON, forcing language models to give only workable answers to hard combo tasks. Such models crack combo tasks in words yet lack any guard that answers truly work, as real use demands. FALCON pairs grammar-locked decoding and a fix layer for rule breaks, and smart best-of-N draws spending compute well. Its train rule weights like pairs by score gaps, giving dense teaching, no hand marks. On seven hard combo tasks, it hit full feasibility while matching or beating top solvers.

Catalog record

CREATORYang Liu, Chuan Zhou, Yancheng Chen, Shuai Zhang, Xixun Lin, Xiaoqing Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE105246
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01090

Browse the full library · About the Tower of Babel