DuoDiscover turns evaluation into agent search, 2x experts on SWE-fficiency
OfirPress · x · 2026-10-15
DuoDiscover tackles the core bottleneck of AI-driven engineering — ideas are cheap to generate but expensive to evaluate — by making evaluation part of what the agent discovers during search. Key results: #1 on NVIDIA's SOL-ExecBench; 2× human experts on SWE-fficiency (work from Jeffrey Ma / Ofir Press's group); 1.92× speedup on vLLM, 1.44× on FreeToken, and 2.68× over production TPU kernels; plus beating the best human- and AI-designed CPU prefetchers. The authors call it a step toward self-maintaining codebases.
Related event: DuoDiscover Tops NVIDIA SOL-ExecBench, Speeds Up TPU Kernels 2.68x(5 posts)→
More from coding & agent
- Microsoft's AutoEnvScaling lets agents build their own RL environments, lifting Qwen3 Terminal-Bench from 41.1 to 53.3 — omarsar0 · 2026-10-15
- tiden_ai verifies AI-generated code against actual requirements, not session tests — Shruti_0810 · 2026-10-15
- With a verification system in place, every new feature is just a prompt away — verrsane · 2026-10-15
- Nobody prompt-caches their voice agent: cache warming is the easiest latency win — TheMoonMidas · 2026-10-15
- Meta's Muse Wonder (Max) lands at #12 in Agent Arena at 2.1x the cost of its predecessor — arena · 2026-10-15
- "If my agent can't browse your site, I won't shop there": Amazon's agent blockade sparks backlash — mimi10v3 · 2026-10-15