Large Language Models (LLMs) achieve great performance by learning effectively from large amounts of data. This can be a problem however when data can become outdated or when data is private. For example:
- LLMs trained last year cannot output accurate information about today's weather.
- LLMs cannot be used to output health information, since that would require training on people's health information.
To avert such limitations and use LLMs in more domains, this paper suggests augmenting the model with a text-based API tools that can access up-to-date or private data. By training the model through iterative self-play, the model can be trained with few labeled examples of tool interaction and achieve better performance in QA tasks than vanilla LLMs.
This paper may be interesting to you if you:
- have private or changing data but want to take advantage of LMs
- want to wrap your APIs with natural language to create a chatbot
Tool-Augmented Language Models (TALM)
Iterative Self-play
Results
TALM is tested on two QA datasets: Natural Questions (NQ) and MathQA.
Conclusion
TODO: Short 1-2 lines
Please read the paper if you want to learn more about:
- how important iterative self-play is to the performance of TALM
- see more examples of TALM's interactions with text APIs
Some other relevant papers that could be interesting to read are: