论文通过model specialization，牺牲了小模型的通用性，提高了小模型在COT推理上的能力。具体而言，论文做出了如下贡献：

• 证明在LM的多维能力间存在复杂的平衡，可以通过损失通用能力来提高数学推理能力；
• 通过动态规划对齐tokenizers。

Fine-tuning ⇒ Specializing

• 1.which foundation models are based on?

• 2.what tokenizers are adopted?

• 3.which datasets are collected specific for “math”?

• 4.what types of pre-processing methods are introduced?

Evaluation

• which datasets are used?

• what evaluation metrics are used?

• 5.other information that you think is important