Thinking Tokens for Language Modeling

🌈 Abstract

The article discusses the limitations of language models in performing complex calculations and reasoning tasks, and proposes the use of "thinking tokens" to enhance the generalization capability of these models.

🙋 Q&A

[01] Thinking Tokens for Language Modeling

1. What is the core idea behind the proposed "thinking tokens"?

• The core idea is to introduce special "thinking tokens" () after each word in a sentence whenever a complex problem is encountered.
• The "thinking tokens" would allow the model to perform more calculations before providing an answer, enabling it to better handle complex tasks that require reasoning.
• This concept has potential in recurrent neural networks due to their ability to perform multiple in-memory operations in a single step.

2. What are the key goals of the proposed approach?

• To enhance the generalization capability of language models and allow them to adapt to more complex tasks.
• To enable the model to decide for itself how much extra time is needed to produce the best answer possible, improving its adaptability and generalization.

3. What are the preliminary results of using "thinking tokens"?

• Experiments show that the use of "thinking tokens" leads to an improvement in the model's performance on sentences that require non-trivial reasoning.
• The introduction of "thinking tokens" is also successful for sentences that include specific numbers or numerical values.

[02] Related Work and Future Work

1. What is the related work mentioned in the article?

• Research on reasoning can be traced back to 1959, and continues to be a focus in the field of theorem proving.
• Large language models are currently being used to learn reasoning from natural language.

2. What are the plans for future work?

• Building on the proof of concept, the researchers plan to extend their research and create a model that can decide for itself how much extra time is needed to produce the best answer possible.
• If successful, this concept could be implemented as a default behavior for language models that encounter complex and computationally demanding tasks.
• The researchers believe that the ability of a model to self-regulate this factor would vastly improve the adaptability and generalization capability of language models in general.