Summary

  • Recent research from Google DeepMind and Stanford University has compared the generalisation capabilities of large language models (LLMs) that have been fine-tuned and those that have undergone in-context learning (ICL).
  • While ICL proved more successful in generalisation tasks, it requires more computing power during inference.
  • Additionally, the researchers proposed a new approach to improve the generalisation capabilities of fine-tuned LLMs by augmenting the learning data using ICL to generate more richly inferred examples.
  • The augmented data can then be used to fine-tune the model, improving its performance on generalisation tasks.
  • The researchers suggested the hybrid approach could lead to more robust and reliable LLM applications, although they admitted it makes the model-tuning process more expensive.

By Ben Dickson

Original Article