DuoDiscover uses proxy evaluators to slash evaluation cost, beats production TPU kernels by 2.68x
CShorten30 · x · 2026-10-15
DuoDiscover tackles a key bottleneck: agents can generate engineering ideas in seconds, but testing them takes hours or days—evaluation cost is what holds AI back in engineering.
- Method: targeted "proxy evaluators" for intermediate steps make evaluation part of what agents discover, dramatically speeding up agentic hillclimbing
- Results: #1 on NVIDIA's SOL-ExecBench; 2x experts on SWE-fficiency; 1.92x on vLLM, 1.44x on FreeToken; 2.68x over production TPU kernels; beats best human and AI-designed CPU prefetchers
- Takeaway: as models improve, efficient evaluation is becoming the bottleneck for test-time-scaling algorithms—folding evaluation into the search loop is a direction worth watching
Related event: DuoDiscover Tops NVIDIA SOL-ExecBench, Speeds Up TPU Kernels 2.68x(7 posts)→
More from coding & agent
- exe.dev Builds Radiant, a Fully Agent-Driven Figma Rival That Ships Real Code — davidcrawshaw · 2026-10-15
- The top skill for coding agents: nail naming, folder structure, and data models first — weswinder · 2026-10-15
- Thud Audio Action Model Claims 10x Faster, 20x Cheaper Than Gemini Live — AnhPhuNguyen1 · 2026-10-15
- Distill a Tiny Model per NPC Instead of Using Frontier LLMs, Says Quintin Pope — QuintinPope5 · 2026-10-15
- Sentry CLI weekly downloads jump from 1.7M to 3M as cloud coding agents pile in — zeeg · 2026-10-15
- Entire Minecraft Crammed Into Elden Ring as an AI-Built Mod — DeryaTR_ · 2026-10-15