The Question Damage Score measures how much LLM reasoning depends on specific context examples versus prior knowledge. By using targeted deletions in linguistic puzzles, the framework classifies reasoning tasks as fragile or robust.
HOW THIS AFFECTS YOU
●
researcherYou can use this diagnostic to determine if models are truly reasoning from context or just relying on training data.