Beyond Scaling: Self-Evolving LLM Agents for Hardware Kernel Optimization via an Experience-Driven Workflow and Experience Graph Memory
3.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.25570.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryKOPE is a framework for hardware kernel optimization using LLM agents. It stores optimization trajectories with correctness and performance feedback in an Experience Graph Memory, then uses Active Context Management to retrieve relevant past experience within a fixed token budget. Reported gains include 1.54x geometric-mean speedup over CANNBot and pass rate rising from 55.2% to 84.6% on a 53-operator suite, with token consumption reduced from 15.9B to 1.113B.
Why it mattersConcrete quantitative result showing that structured experience memory lets an agent learn across runs while the underlying model stays fixed. The token-budget and pass-rate numbers are directly comparable to existing baselines.
Cited by
No citations on record.