ACLJan, 2024

大型语言模型的过度推理和冗余计算

TL;DRLLMs tend to generate lengthy and unnecessary calculations on the math QA dataset GSM8K-Zero, even though the questions can be answered without any calculations.