Directional Stimulus: Definition and Examples

Directional Stimulus Prompting is a method introduced by researchers from the University of Hong Kong in 2023. It is based on the idea of adding a 'directional stimulus' — a small hint, keyword, or prompt — to the prompt in order to guide the model's reasoning in the right direction. Unlike a direct instruction, the stimulus acts as a subtle signal that pushes the LLM to explore a particular reasoning path.

In practice, the directional stimulus can take many forms: a relevant keyword, a draft outline, a targeted contextual element, or a specific constraint. The goal is to reduce prompt ambiguity and channel the model's generation toward more precise and relevant outputs, without restricting its creativity.

This technique differs from classic approaches like few-shot prompting because it does not require providing complete examples. Instead, it relies on minimal but strategically chosen hints. In the original article, the researchers even proposed training a small auxiliary model (a 'policy model') capable of automatically generating these optimal stimuli for a main LLM.

Directional Stimulus is particularly effective for summarization, dialogue generation, and reasoning tasks, where a simple hint can significantly improve the quality and relevance of responses. It is an elegant approach that exploits the ability of LLMs to pick up on weak signals to produce significantly better results.