paper · First submission: April 19, 2024
The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
Training conditional obedience: lower-priority requests can be useful when they do not conflict.
Eric Wallace, Kai Xiao, Reimar Leike, Lilian Weng, Johannes Heidecke and Alex Beutel train GPT-3.5 Turbo on hierarchical instruction following. Compatible lower-priority requests remain useful; conflicting ones should be ignored. The study reports improved robustness alongside regressions on two over-refusal evaluations (§4).[1]
Read Instruction Hierarchy for an original example. Compare system-level design to distinguish model behavior from application permissions.
Historical context
Original arXiv v1 submitted April 19, 2024. Selected hierarchy, training, evaluation and conclusion checked; rendered p.9 inspected. Complete datasets and replication were not audited. The authors explicitly expect remaining vulnerability to powerful adversarial attacks; this is not a proof of universal protection.[1]
Sources
Pages that link here
Last updated 2026-10-12