Google's DuoDiscover Surpasses Human Experts on SWE-fficiency and Beats ArchAgent v2 on Chip Design
google · hf · 2026-10-19
Google, UC Berkeley and Stanford present DuoDiscover, a search system for autonomous discovery in real-world engineering where a single true evaluation can take hours.
Three key mechanisms:
- Each coding-agent worker co-discovers an optimization target and a fast tailored proxy evaluator ("yardstick"), hill-climbing the yardstick instead of the true objective
- Fidelity tests continually recheck each yardstick's ranking and invalidate any that drifts, guarding against proxy gaming
- The fleet is bottom-up with no central decomposition; a lightweight orchestrator spawns/terminates agents, briefs running ones, and seeds new agents with confirmed work
Results:
- Doubles the average speedup of human experts on SWE-fficiency
- SOL-Exec: ranks first on 121 of 235 GPU/TPU kernels, top-3 on 212
- 1.92x gains on heavily optimized vLLM and 1.38x on SQLite
- On a CPU microarchitecture task with evaluations up to 28 hours, surpasses AlphaEvolve-based ArchAgent v2 (1.65x its gain in 16.7 days vs 40+) and the best human design
More from coding & agent
- Dev ships a playable Tesla Roadster game in a weekend using Opus 5.5 — edwinarbus · 2026-10-19
- Musk amplifies Grok Bot feature list: own computer, email, and a team of bots — elonmusk · 2026-10-19
- HIVE team on agent access layers: approvals must bind exact arguments and policy versions — usehive · 2026-10-19
- PoA: An Experimental Network Where AI Agents Earn Coins by Writing Original Work — adamamcbride · 2026-10-19
- Next Token podcast: 'If AI gets something to 60 points, it's worth zero' — op7418 · 2026-10-19
- Running GLM on 4x DGX Spark locally to build a 3D Three.js portfolio site — 9r4n4y · 2026-10-19