DuoDiscover has agents design their own evaluators, beating production TPU kernels by 2.68x
CShorten30 · x · 2026-10-15
DuoDiscover: scaling autonomous discovery past the evaluation bottleneck
Agents can generate candidate solutions in seconds, but a single true evaluation (full benchmark or serving trace) can take hours, capping any search at a handful of real evaluations.
The paper introduces three mechanisms:
- Co-discovered yardsticks: each coding agent jointly discovers an optimization target and a fast, tailored proxy evaluator, then hill-climbs the yardstick instead of the true objective.
- Fidelity tests: rankings against the true objective are continually rechecked; drifting yardsticks are invalidated to prevent proxy gaming.
- Bottom-up agent fleet: no central decomposition; a lightweight orchestrator just spawns, terminates, and briefs agents.
Results: #1 on NVIDIA's SOL-ExecBench; 2x experts on SWE-fficiency; 1.92x on vLLM and 1.44x on FreeToken; 2.68x over production TPU kernels; beats the best human- and AI-designed CPU prefetchers.
The 23-author team includes Ion Stoica and Nanyun Peng; code is coming soon.
Related event: DuoDiscover Makes Evaluation Part of the Search, Boosting TPU Kernels 2.68x(3 posts)→
More from coding & agent
- Zenzap raises $44M and launches AI agents in its workplace chat app — testingcatalog · 2026-10-15
- LangChain Deep Agents demo: drop a CSV in Slack, get a slide deck back — Hacubu · 2026-10-15
- 'Slop' perception about to reverse: soon not using AI may signal sloppy coding — mark_k · 2026-10-15
- Workday and Anthropic connect HR and finance agents via MCP — emmanuelvivier · 2026-10-15
- Shopify opens its billion-listing catalog to any AI agent with one-click Shop Pay checkout — emmanuelvivier · 2026-10-15
- DuoDiscover turns evaluation into agent search, 2x experts on SWE-fficiency — OfirPress · 2026-10-15