MoST Mines Optimization Strategies From Textbooks and Web Pages, Not Just Commits — Up to 717% Speedups on Real Projects

arXiv 2607.20353·medium signal

Current strategy-guided code optimizers only formalize rules from historical optimization commits, and only for the language the commit came from. MoST represents items from any knowledge source as uniform evidence objects, clusters them cross-source and cross-scenario with a self-balanced weighted algorithm, and transfers strategies to a target language when needed. On 151 C/C++, 150 Python, and 50 Rust optimization tasks it produces 24.44%–180.00% more patches identical to developer patches than SemOpt; on 15 real projects it delivers 4.44%–258.17% average and up to 717.42% maximum performance-test improvements, beating both SemOpt and Codex.