Which Prompts Make The Difference? Data Prioritization For Efficient Human LLM Evaluation

Large language models (LLMs) have produced notable breakthroughs in downstream performance [61; 11; 19; 62; 91; 49; 8; 78], but have also introduced new challenges in model evaluation. The success of LLMs has initiated a fundamental paradigm shift away from small specialized models designed for single tasks to universal models expected to perform well across a wide range of tasks. This shift has also posed an existential challenge for evaluation, with a need to move away from solely task-specific automatic metrics of evaluation and increasing reliance on human evaluation. While automatic metrics offer a degree of objectivity and reproducibility, alongside the benefits of speed and cost-effectiveness, they often fall short in fully capturing the complexities and nuances of natural language [48; 68]. Moreover, automatic metrics often rely on auxiliary models which introduce potential points of failure and unexpected challenges over time [58]. For example, reference-based metrics such as BLEU [54] and ROUGE [45] are usually poor indicators of human judgment, as they emphasize lexical overlap and struggle to account for the diverse expressions inherent in semantic representation [34; 84; 9].