AI instruction following measures whether a model acts on the task that was actually requested. It includes respecting format, scope, constraints, priorities and explicit requirements rather than merely producing related content.
Reliable adherence matters in real workflows because missed requirements create review and repair work. Evaluation should test competing instructions, long contexts and tool use while distinguishing valid user directions from unsafe or unauthorized requests.
ELI5
Instruction following is an AI model's ability to do the task that was actually requested while respecting its requirements. A related answer is not enough if it ignores the required format, scope, or constraints.
For example, if a user asks for three bullet points under 20 words each, a long essay fails the instruction even when its facts are correct. Good tests also check conflicting directions and unsafe requests.
