{"@context":"https://schema.org","@type":"CreativeWork","@id":"https://forgecascade.org/public/capsules/6aa83c01-1f96-4499-bfcd-8336349d933e","identifier":"6aa83c01-1f96-4499-bfcd-8336349d933e","url":"https://forgecascade.org/public/capsules/6aa83c01-1f96-4499-bfcd-8336349d933e","name":"What Makes In-Context Learning Demonstrations Work?","text":"Min, Lyu, Holtzman, Artetxe, Lewis, Hajishirzi, and Zettlemoyer analyze which parts of few-shot demonstrations matter for in-context learning. The paper reports that correct input-label mappings are often less important than demonstration format, label space, and input distribution cues across tested classification settings. Use this as a source-backed reference for prompt demonstration effects, not as a universal claim that labels never matter.\n\nSources:\n- https://arxiv.org/abs/2202.12837","keywords":["in-context-learning","demonstrations","prompt-format","label-space"],"about":[],"citation":[],"isPartOf":{"@type":"Dataset","name":"Forge Cascade Knowledge Graph","url":"https://forgecascade.org"},"publisher":{"@type":"Organization","name":"Forge Cascade","url":"https://forgecascade.org"},"dateCreated":"2026-04-11T06:57:23.088633Z","dateModified":"2026-06-19T01:57:15.363000Z","isBasedOn":"https://arxiv.org/abs/2202.12837","additionalProperty":[{"@type":"PropertyValue","name":"trust_level","value":95},{"@type":"PropertyValue","name":"verification_status","value":"sources_verified"},{"@type":"PropertyValue","name":"provenance_status","value":"valid"},{"@type":"PropertyValue","name":"evidence_level","value":"primary_source"},{"@type":"PropertyValue","name":"content_hash","value":"396a3b0bb233d0dd9b2571f8286e4a278847b986f75bc2da5b514727bf2bca3f"}]}