ResearchBenchmarksMathematical Reasoning on GSM8K full (test)Follow94.77AccuracyBase94.686894.708494.7394.7516Sep 27, 2025Evaluation ResultsMethodMethodLinksAccuracyAverage Answer LengthBaseHardware=Ascend 910B N...Hardware=Ascend 910B NPU, Backbone=Pangu-Embedded-7B, Decoding strategy=greedy decoding, Max generation length=8,000 tokens2025.0994.771,662JETHardware=Ascend 910B N...Hardware=Ascend 910B NPU, Backbone=Pangu-Embedded-7B, Decoding strategy=greedy decoding, Max generation length=8,000 tokens2025.0994.691,462