Instruction following
A model’s ability to actually do what you tell it, not just continue your text.
Think of it like
Like the difference between a parrot that echoes you and an assistant that hears “summarize this” and summarizes.
Example
You write “Reply in exactly two sentences, no bullet points,” and a well-tuned model obeys both constraints instead of ignoring them.
How it actually works
Base models just predict likely text; instruction following is a learned behavior, mostly from instruction tuning and RLHF, that makes them treat prompts as commands. It’s never perfect — models miss constraints, especially several at once — which is why clear, checkable instructions still matter.
For product teams
The trait that turns a raw model into something a non-expert can actually direct.
For engineers
Learned behavior (via SFT/RLHF) to condition outputs on directives rather than mere continuation; degrades with constraint count.
Related
- RLHF — The training that instills it.
- Base model — What a raw model does instead.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome