Thinking Tokens for Language Modeling

TL;DR

This paper introduces 'thinking tokens', a novel method that enables language models to perform complex calculations more effectively by mimicking human problem-solving processes.

Contribution

The paper proposes the use of special 'thinking tokens' to improve language models' ability to handle complex reasoning tasks, enhancing their generalization capabilities.

Findings

Thinking tokens improve calculation accuracy in language models.

Models with thinking tokens outperform baseline models on complex reasoning tasks.

Enhanced reasoning ability observed across multiple benchmarks.

Abstract

How much is 56 times 37? Language models often make mistakes in these types of difficult calculations. This is usually explained by their inability to perform complex reasoning. Since language models rely on large training sets and great memorization capability, naturally they are not equipped to run complex calculations. However, one can argue that humans also cannot perform this calculation immediately and require a considerable amount of time to construct the solution. In order to enhance the generalization capability of language models, and as a parallel to human behavior, we propose to use special 'thinking tokens' which allow the model to perform much more calculations whenever a complex problem is encountered.