Prompt Engineering

Prompt Engineering

Prompt-engineering is taking the embedding representation of LLMs and using them to help reduce one of the difficulties in NLG, namely the inability to guide the models to generate more useable language. This can be implemented and achieved using different architectural implementations, including hard prompt tuning, soft prompt tuning, using Reinforcement Learning from Human Feedback (RLHF) among others. Rather than a static defined outcome (such as a set of labels), the user has the ability to create prompts that contain what the model should expect as input, how it should think about this input, and the output representations (see Figure 3), the latter of which being perhaps the more important development from the generation perspective. Representations are not only limited to text, they can be structured formats or code as whatever provide as long as long as it can be transformed into a representation the model can process.

Figure 3: Prompt Engineering Training

From a user perspective, this does not require deep technical knowledge of the underlying mechanics of the model, as much of the sorcery is performed underneath (such as the textual conversion to embeddings). Rather the approach allows the focus to predominantly be on understanding how to provide optimal examples of what the model would be expected to process and produce to yield the optimal outcomes. The two aspects of being able to provide specific inputs and information paired with the desired outputs in a shortened loop to improve the model heavily mitigates many of the deficiencies within traditional NLP models as well as initial embedding models. This is one reason it has gained such wide popularity as well as possessing a lower barrier to entry for those not technically knowledgable in the the field. Currently, assuming a good use case fit, prompt engineering is one of the methods by which an individual can provide the most localized control over the desired behavior of a model to achieve a desired outcome.

Related articles