In the world of LLMs, the focus often lies on perfecting single prompt inputs for In-Context Learning. However, using multiple prompt inputs in tandem might further elevate LLM performance. Introducing In-Context Sampling (ICS), a low-resource LLM prompt-engineering technique aimed at optimizing multiple prompt constructions to boost prediction results. This approach has been tested across several state-of-the-art LLMs and NLI datasets, revealing its potency in enhancing LLM’s predictive abilities and certainty.

Significant Takeaways:

• Insight into the under-explored area of utilizing multiple prompt inputs collectively.
• The proposed ICS technique’s effectiveness in improving prediction accuracy and confidence.
• Extensive testing providing evidence for ICS’s benefits.
• An ablation study pointing towards a diversity-based ICS strategy.

As the researchers shed light on a promising new strategy within the domain of AI, they push the boundaries of what LLMs can achieve with fewer resources. This work guides future research into deeper exploration and optimization of prompt inputs and their collective use. Further details are available here