Self-Improving In-Context Learning

In three linesMethod to improve in-context learning by optimizing continuous embeddings of few-shot prompts at test time. Uses model log-probabilities on demonstrated outputs as self-supervised signal, requiring no token generation or fine-tuning. Improves performance on both classification and free-form generation tasks.

Summary generated by Claude — human-verified