DuoDiscover

TL;DR

DuoDiscover has coding agents discover both an optimization target and a cheap proxy evaluator, then repeatedly checks whether that proxy still tracks the expensive real objective. See Agents (LLM Agents). (source)

Authors & Org

Shubham Agarwal and coauthors; affiliations are not established by the abstract record. First submitted 2026-10-12. The snapshot preserves the complete author list and equal-contribution note. (source)

Method

Workers independently propose optimization targets and fast evaluators called yardsticks. They optimize against those yardsticks; fidelity tests recheck their rankings and invalidate drifting proxies. A lightweight orchestrator starts and stops workers and shares confirmed findings, without centrally decomposing the problem. (source)

Results

The authors evaluate 2 repository-level benchmarks and 4 GPU/TPU kernel suites. On SWE-fficiency, they report double the average human-expert speedup. On SOL-Exec, the system ranks first on 121 of 235 kernels and in the top three on 212. Reported gains over base code are 1.92× for vLLM and 1.38× for SQLite. (source)

For CPU microarchitecture optimization with evaluations taking up to 28 hours, the authors report 1.65× ArchAgent v2's gain in 16.7 days, versus more than 40 days for that comparator. This is a multiplier on the comparator's gain, not an absolute hardware speedup. (source)

Significance

The interesting unit of Eval Harness Configuration is a proxy plus an ongoing check of its fidelity. Cheap evaluation only helps if optimizing it continues to improve the actual objective. (source)

Open Questions

The abstract does not give total agent cost, fidelity-test overhead or failure rates for invalidated proxies. Full paper and code were not reviewed; the reported improvements are not independently replicated here. (source)

Cite

“DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck.” arXiv:2610.17184. This note uses the abstract and metadata in the snapshot.