RAPT: Retrieval-Augmented Post-hoc Thresholding for Multi-Label Classification

In three linesRAPT is a retrieval-augmented post-hoc thresholding wrapper improving label set selection in multi-label classification without retraining. Applied to metric learners and fine-tuned transformers, RAPT achieves 0.87 Macro-F1 on industrial data, outperforming static baselines and few-shot LLMs (K=5) with 115x less inference time.

Summary generated by Claude — human-verified