Adaptive Prompting in the Metaverse: An Iterative Prompt Optimization Framework for Enhancing LLM Performance Across Diverse Tasks

Metaverse applications increasingly demand dynamic and context-sensitive language interactions, presenting substantial challenges to conventional prompt engineering methodologies. This paper introduces AdaptiPrompt, an iterative optimisation framework designed to refine prompts for large language models automatically, thereby enhancing performance in diverse fields that range from college-level scientific discourse to nuanced ethical analysis. The proposed approach integrates a dataset generation pipeline—where user instructions undergo ongoing revisions informed by teacher feedback—and a specialised Prompt Optimiser that generates precisely tailored prompts in real time. Empirical evaluations on key benchmarks, including MMLU, MMLU-Pro, and GSM8K, yield competitive outcomes, especially within scientific and policy-driven tasks. However, Chain-of-Thought prompting retains a minor advantage when addressing formal mathematics. These results indicate that adaptive prompt optimisation, as embodied by AdaptiPrompt, holds significant potential for emerging metaverse scenarios that call for context-aware and low-latency responsiveness. Such adaptability could be pivotal in ensuring that large language models function effectively and efficiently across increasingly multifaceted use cases in virtual environments.