Which of the following describes the scenario where an LLM is embedded 'As-is' into an application frame?

The term 'As-is' in software engineering and AI model deployment refers to the integration of a component without any alterations to its core structure or pre-trained state. In the context of an LLM, this means embedding the model directly into an application and utilizing its existing, general-purpose capabilities through API calls. This approach, often facilitated by zero-shot or few-shot prompting, leverages the model's out-of-the-box functionality without undergoing task-specific fine-tuning or customization. The application's frame essentially acts as a user interface for the unmodified, pre-trained LLM.

B. This describes replacing the model entirely, which is the opposite of embedding the current one 'as-is'.

C. This involves customization and fine-tuning, which fundamentally alters the model from its 'as-is' state to suit specific needs.

D. This defines an architectural placement (backend processing) but does not specify the state of the LLM; a customized or an 'as-is' model could be used this way.

1. Zhao

W. X.

et al. (2023). A Survey of Large Language Models. arXiv preprint arXiv:2303.18223. Section 4.2

"Tuning-free Prompting

" describes methods like zero-shot and few-shot learning where the pre-trained model's parameters are not updated

which is synonymous with using the LLM 'as-is'. (https://doi.org/10.48550/arXiv.2303.18223)

2. Yang

J.

et al. (2023). Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond. arXiv preprint arXiv:2304.13712. Section 3

"Paradigms of LLM-based Applications

" contrasts approaches that modify the LLM (e.g.

fine-tuning) with those that treat the LLM as a "frozen" or unchangeable component

which corresponds to an 'as-is' integration. (https://doi.org/10.48550/arXiv.2304.13712)

3. Stanford University. (2023). CS224N: Natural Language Processing with Deep Learning. Lecture 10

"Prompting

Instruction Finetuning

and RLHF." The lecture distinguishes between using a base pre-trained model directly via prompting (the 'as-is' scenario) and adapting the model's weights through fine-tuning for specific tasks.