ResearchBenchmarksBig-Bench Hard Performance and Efficiency on BBHFollow0.476Normalized Log AccuracyT-Free0.471840.472920.4740.47508Mar 16, 2026Evaluation ResultsMethodMethodLinksNormalized Log AccuracyBytes per Sequence PositionT-Freeshots=3-shotshots=3-shot2026.030.4764.67HATifiedshots=3-shotshots=3-shot2026.030.4724.67Llamashots=3-shotshots=3-shot2026.030.4723.79