מושגי ליבה
Small models can efficiently learn Chain-of-Thought (CoT) through AS-ES learning, maximizing existing data without additional augmentation.
תקציר
AS-ES learning introduces a new training paradigm for small models to improve CoT learning efficiency. The method involves segmenting CoT data into Extractive Segments (ES) and Abstractive Segments (AS). This approach enhances logical reasoning capabilities without altering the model or requiring extra data. Experimental results show improved performance on tasks like Math Word Problems and PET summarization. The study explores the impact of segmentation strategies, model sizes, and hyperparameters on the effectiveness of AS-ES learning.
סטטיסטיקה
Weng earns $12 an hour for babysitting.
Yesterday, she just did 50 minutes of babysitting.
How much did she earn? Let's think step by step.
Math Word Problem
Answer
Weng earns $12 per hour,
which means she earns $1 per 5 minutes.
She babysat for 50 minutes,
which means she earned 50/5 = 10 dollars,
The answer is 10.
ציטוטים
"Existing methods often simply generate and incorporate more data from LLMs and fail to note the importance of efficiently utilizing existing CoT data."
"We introduce AS-ES learning, a novel data-efficient training paradigm that maximizes the intrinsic value of existing CoT data."
"We provide a theoretical foundation for the efficacy of AS-ES learning, offering insights into the underlying dynamics of CoT."