IHEval shows that current language models often follow lower-priority instructions over system messages, and simple prompting does not fix the problem.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
IHEval: Evaluating Language Models on Following the Instruction Hierarchy
IHEval shows that current language models often follow lower-priority instructions over system messages, and simple prompting does not fix the problem.