ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends

Shengjun Kris Dong

Dima Nikiforov

Sahil Modi

Sana Damani

Edward C Lin

Siva Kumar Sastry Hari

Christos Kozyrakis Stanford

Preprint, 2026


Abstract

ModelBlaster uses memory-augmented in-context reinforcement learning to optimize kernels and whole models across heterogeneous accelerators. A persistent knowledge base lets agents reuse optimization experience across CUDA, OpenCL, and RISC-V vector backends.