IHEval: Evaluating Language Models on Following the Instruction Hierarchy
This addresses a critical gap in evaluating language models for safety and consistency in following instruction priorities, though it is incremental as it focuses on benchmarking rather than proposing new methods.
The paper tackles the problem of evaluating language models' ability to follow the instruction hierarchy, which prioritizes system messages over user messages, conversation history, and tool outputs for consistent and safe behavior. It introduces IHEval, a benchmark of 3,538 examples across nine tasks, showing that popular models struggle with conflicting instructions, with the best open-source model achieving only 48% accuracy in such cases.
The instruction hierarchy, which establishes a priority order from system messages to user messages, conversation history, and tool outputs, is essential for ensuring consistent and safe behavior in language models (LMs). Despite its importance, this topic receives limited attention, and there is a lack of comprehensive benchmarks for evaluating models' ability to follow the instruction hierarchy. We bridge this gap by introducing IHEval, a novel benchmark comprising 3,538 examples across nine tasks, covering cases where instructions in different priorities either align or conflict. Our evaluation of popular LMs highlights their struggle to recognize instruction priorities. All evaluated models experience a sharp performance decline when facing conflicting instructions, compared to their original instruction-following performance. Moreover, the most competitive open-source model only achieves 48% accuracy in resolving such conflicts. Our results underscore the need for targeted optimization in the future development of LMs.