ResearchBenchmarksReward rollout alignment on 10-task benchmark T1: Folding ShortsFollow0.889Rollout ρSARM2-0.266440.033530.33350.63347Jun 9, 2026Evaluation ResultsMethodMethodLinksRollout ρSARM2training_subset=S1training_subset=S12026.060.889SARM2standard_full=truestandard_full=true2026.060.833RobometerFine-tuning=trueFine-tuning=true2026.060.667SARM2Multi-Goal=excludedMulti-Goal=excluded2026.060.611SARM2Spectral Embedding=exc...Spectral Embedding=excluded2026.060.556RobometerFine-tuning=falseFine-tuning=false2026.060.222RW2026.060.167TR2026.06-0.222