ArchAgent v2 Wins DPC4: Divide-and-Conquer Search Beats Human Designs

rohanpaul_ai · x · 2026-08-24

Google DeepMind and UC Berkeley released ArchAgent v2, which defeated the human-designed champion DPC4 in the Data Prefetching Championship. The agent employed a divide-and-conquer strategy, optimizing cache levels (L1D/L2/LLC) in stages while strictly adhering to storage budget constraints, followed by joint refinement. The final design achieved a 4.6% improvement over BertiGO on low-bandwidth single-core systems (vs. 2.6% baseline) and a 3.8% overall IPC gain. The study highlights that structured search and real engineering constraints are more critical than model strength in AI discovery tasks.

Original post →

More from Research

Research channel →