.png){kind=icon}
.png){kind=icon}
Return
Behavior and representation in open-weight Large Language Models for combinatorial optimization: From feature extraction to algorithm selection
.png){kind=icon}
OA
AI
DOI：10.1016/j.cor.2026.107603.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•
A unified framework to evaluate what large language models (LLMs) learn about combinatorial optimization problems (COPs) through direct querying and probing.
•
Validation across five open-weight LLMs from 3B to 120B parameters.
•
Analysis of the scale-dependent effect of chain-of-thought prompting on feature extraction performance.
•
Direct querying and probing capture complementary aspects of feature information, corresponding to explicit recoverability and latent decodability.
•
LLM-derived embeddings match handcrafted features in algorithm selection, with a trade-off between per-instance inference cost and feature-engineering effort.
Keywords:
Combinatorial optimization
Large Language Models
Direct querying
Probing
Feature extraction
Algorithm selection
Journal
C
IF:
4.3
Papers:
201
Citations:
0