ResearchBenchmarksDiverse Reasoning on BIG-bench HardDiverse Reasoning on BIG-bench HardFollow13.99BBH QA ScoreEngram2.79965.70488.6111.5152Jun 6, 2026Evaluation resultsMethodMethodLinksBBH QA ScoreBBH Dyck ScoreBBH CS ScoreBBH Operator ScoreEngramLayers (L)=18Layers (L)=182026.0613.9914.840.9812.38TN-gramLayers (L)=18Layers (L)=182026.0612.3215.241.3616.19Raw GPTLayers (L)=18Layers (L)=182026.069.0516.939.7715.71TN-gramLayers (L)=9Layers (L)=92026.066.2913.937.4212.38EngramLayers (L)=9Layers (L)=92026.064.4613.938.9411.9Raw GPTLayers (L)=9Layers (L)=92026.063.2313.438.4113.33