Easier to Judge than to Find: Predicting In-Context Learning Success for Demonstration Selection

In three linesDiSP, a demonstration selection framework for in-context learning, predicts whether a query-context pair will succeed rather than searching for optimal context. On 5 classification datasets with Llama 3-8B and Qwen 2.5-7B, DiSP improves accuracy by 3.4% and achieves 23× end-to-end speedup.

Summary generated by Claude — human-verified