ResearchBenchmarksReasoning on HellaSwag (Accuracy and Compression)Follow82.6Normalized Log AccuracyT-Free79.16880.05980.9581.841Mar 16, 2026Evaluation ResultsMethodMethodLinksNormalized Log AccuracyCompression (Bytes/Position)T-Freeevaluation_protocol=10...evaluation_protocol=10-shot2026.0382.65.34Llamaevaluation_protocol=10...evaluation_protocol=10-shot2026.0381.84.66HATifiedevaluation_protocol=10...evaluation_protocol=10-shot2026.0379.35.34