Subset Selection via Rank Correlation Prediction for Data-Efficient LLM Evaluation (SubLlME) is a novel evaluation method proposed by HP Labs and other teams in July 2025. It aims to achieve efficient and accurate model performance evaluation through rank correlation prediction without the need for full evaluation.SubLIME: Subset Selection via Rank Correlation Prediction for Data-Efficient LLM Evaluation", which won the ACL 25 Best Theme Paper Award.

SubLIME draws on the evaluation strategy of the Olympic Mathematical Competition. By intelligently selecting a small but representative subset, it predicts the relative performance of models in the full evaluation, thereby significantly reducing the evaluation cost (reducing 80%–99%) while maintaining highly consistent model ranking results.