DuoDiscover Tackles the Evaluation Bottleneck, Claims 2.68x TPU Kernel Gains
stanfordnlp · x · 2026-10-15
- Stanford, Berkeley and Google Cloud researchers published DuoDiscover, targeting the biggest bottleneck in autonomous scientific discovery: evaluation cost — agents can generate ideas in seconds, but testing them takes hours or days.
- Key idea: make evaluation part of what the agents themselves discover, rather than a fixed external verification step.
- 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; CPU prefetcher designs beat both the best human- and AI-designed ones.
Related event: DuoDiscover Tops NVIDIA SOL-ExecBench, Speeds Up TPU Kernels 2.68x(7 posts)→
More from coding & agent
- Claude Code ships 278 changes in 2 days: 7 updates worth knowing — minchoi · 2026-10-15
- Video edited entirely by Claude Opus 4.5 High using free self-made skills, no editing software — chongdashu · 2026-10-15
- How to Connect Grokbot to Claude Code, Codex or Any AI Agent via HQ Messages — jacob_posel · 2026-10-15
- Stanford's CausalSmith automates causal inference research with Lean-verified proofs — StanfordAILab · 2026-10-15
- ASENA: a coding agent writes robot programs on the fly, no training needed — zhengyiluo · 2026-10-15
- Thorsten Ball: no PRs, no code review — agents push to main and self-verify — kieranklaassen · 2026-10-15