Instruction Following

Instruction following is an LLM behavior where the model treats parts of the input (often system/developer/user messages in a prompt) as instructions and generates outputs that aim to satisfy those constraints.

Details

Instruction following is typically strengthened during post-training via instruction fine-tuning and preference-based methods (for example RLHF), and it strongly influences how "chat" models behave across multi-turn conversations. It is not the same as task capability: a model can understand an instruction yet fail to complete it due to missing knowledge, weak reasoning, or missing tools.

Because instruction following is a trained behavior that the model applies broadly across its input, it can be exploited when untrusted content is treated as instruction - for example a malicious instruction embedded in retrieved content (see prompt injection).

Examples

• "Answer in JSON with keys ...", "Be brief", "Use the provided context only", "Do not reveal secrets".

Synonyms

instruction adherence, instruction compliance