DuoDiscover scales autonomous discovery past the evaluation bottleneck

What the research found

DuoDiscover is an arXiv search system where coding-agent workers optimize tailored proxy evaluators, reporting 2x human speedup on SWE-fficiency and 1.65x ArchAgent v2 gain on CPU microarchitecture in 16.7 days; no code or weights linked.

Claims and how well the evidence supports them

1. DuoDiscover doubles the average speedup of human experts on SWE-fficiency 98%
2. DuoDiscover achieves 1.92x and 1.38x gains over base code on vLLM and SQLite 99%
3. CPU microarchitecture task with evaluations up to 28 hours completed in 16.7 days versus more than 40 days 98%
4. DuoDiscover ranks first on 121 of 235 kernels on SOL-Exec 99%

Evidence

• huggingface.co https://huggingface.co/papers/2610.17184

Actionability rated low.