Submodular Context Partitioning and Compression for In-Context Learning

Sub‑CP is a submodular, block-aware framework for selecting context examples in in‑context learning. It introduces a diversity–coherence spectrum controlled via four partition strategies—Global Diverse, Global–Local Diverse, Local Diverse, and Local Coherent—to balance global coverage and local structure. The framework supports offline precomputation for scalable inference and integrates seamlessly as a pre-processing step in ICL pipelines. Experiments incorporating Sub‑CP into DENSE, ICAE, and CEPE show substantial improvements on TREC, SST‑2/5, MR, and AG News tasks, demonstrating its effectiveness across diverse datasets.