ResearchBenchmarksContinuous-reward optimization on KernelBench (val)Continuous-reward optimization on KernelBench (val)Follow1.089MLP SpeedupEvolver0.987081.013541.041.06646May 11, 2026Evaluation resultsMethodMethodLinksMLP SpeedupShuffleNet SpeedupGRU SpeedupEvolverRefinement iterations...Refinement iterations (R)=22026.051.0891.2182.226No skill2026.051.0271.1171.326EvolverRefinement iterations...Refinement iterations (R)=12026.050.9911.4372.185