KOPE Cuts Kernel-Optimization Token Spend From 15.9B to 1.113B While Raising Pass Rate 60.0% to 84.6%

arXiv 2608.25570·medium signalXBlueskyLinkedInCopy link

Chen et al. store hardware kernel optimization trajectories, with correctness and performance feedback, in an Experience Graph Memory that preserves decision order, observed outcomes, and abandoned branches, then retrieve from it under a fixed token budget via Active Context Management and Injection. Under the same GLM-5.2 setting, KOPE's geometric-mean per-operator speedup is 1.54x that of the strongest baseline, CANNBot. The 53-operator ablation is the striking part: active context management alone lifts pass rate from 60.0% to 84.6% while cutting token consumption from 15.9B to 1.113B versus letting the agent build its own context, and enabling the graph memory lifts full-suite pass rate from 55.2% to 84.6%.