Instruction Following
The model’s ability to actually obey what you asked.
How faithfully a model respects explicit constraints — format, length, prohibitions, ordering. It degrades as instructions accumulate or conflict.
Every added instruction competes with the others. Past a handful of hard constraints, models start dropping some. Keep the must-haves few, explicit and near the end.
Treat constraints as a budget. Each hard requirement competes with the others, and past a handful, the model silently chooses which to honour. The discipline is to keep must-haves few, explicit and near the end, and to verify them with tests rather than assuming compliance.
The forgotten constraint
A prompt with eighteen rules reliably drops two of them, and not the same two.
Move the two that actually matter into a post-processing check rather than trusting the prompt, and compress the rest. Deterministic validation on the output is the honest way to enforce a hard constraint; a sentence in the prompt is a request.
Conflict detection
Instructions say both "be thorough" and "keep it under a page".
The model resolves the conflict differently across runs, which looks like flakiness. Resolving it explicitly — depth in the analysis, brevity in the summary — turns an ambiguous instruction into a specific one and the variance drops.
- paperFinetuned language models are zero-shot learners ↗
Instruction tuning: training on many tasks phrased as instructions improved results on tasks the model had not been trained on.
- paperTraining language models to follow instructions with human feedback ↗
InstructGPT: supervised fine-tuning on demonstrations, then a preference model, then RLHF. Also the result that raters preferred a much smaller tuned model to raw GPT-3.
- docsPrompting best practices ↗
The current house guidance on explicit instructions, output constraints, examples, XML structure and role prompting.
Prompt Engineering
The craft of writing instructions a model reliably follows.
System Prompt · System message
The standing instructions the user does not see.
Instruction Tuning · SFT
The stage that turns a text predictor into something that follows requests.
Context Window
The hard ceiling on how much the model can see at once.