Summarizing https://arxiv.org/pdf/2305.05364

Here's my try:

The paper presents a new approach to expand the capabilities of large pre-trained language models (LLMs) by embedding them within algorithms or programs. The authors provide an example of evidence-supported question-answering and show that their method can improve performance without fine-tuning. They also discuss the advantages and disadvantages of this approach compared to traditional methods.

In addition, the authors highlight some limitations of previous work on in-context learning for LLMs, including the need to prompt engineer multiple algorithm steps within one prompt, which can lead to interference between steps.

Another drawback of training on random internet text is that such raw LLMs also struggle to be conversational, follow instructions, use tools, or interact with an environment. Although finetuning on examples of such behavior has been shown to improve the model's ability (Ouyang et al., 2022; Wei et al., 2022a; Gupta et al., 2022), such an approach is difficult to scale due to the lack of such data and may not generalize systematically due to the possibility of leveraging algorithmic short-cuts (Liu et al., 2021).

The authors propose a new approach that embeds LLMs within algorithms or programs, allowing them to learn from the algorithm's input/output pairs. This approach can be more efficient than finetuning on examples of conversational behavior, as it allows the model to learn from a wider range of inputs and outputs, including those that are not human-generated. The authors also highlight the potential for this approach to improve the robustness of LLMs by providing them with a better understanding of the context in which they will be used.