Measuring Pragmatic Influence in Large Language Model Instructions
A new study quantifies how 13 framing strategies across 400 prompts can shift AI model prioritization.
A team of researchers from Carnegie Mellon University and the University of Edinburgh has published a groundbreaking paper titled 'Measuring Pragmatic Influence in Large Language Model Instructions.' The study tackles a well-known but poorly quantified phenomenon: how the framing of a prompt—phrases like 'This is urgent' or 'As your supervisor'—can systematically bias an LLM's response without changing the core task. The researchers argue that this 'pragmatic framing' has been exploited for optimization or seen as a security flaw, but has never been treated as a measurable property of instruction-following systems. Their work establishes it as a predictable factor that developers must account for.
The core of their contribution is a three-part framework. First, they decompose instructions to isolate framing cues from task specification. Second, they introduce a comprehensive taxonomy that organizes 400 specific instantiations into 13 distinct framing strategies (e.g., appeals to authority, urgency, morality) across four mechanism clusters. Third, they use a 'priority-based measurement' to quantify how these frames shift a model's directive prioritization. Testing across five LLMs from different families and sizes, they found that framing mechanisms cause consistent and structured deviations from baseline impartiality, systematically favoring the framed directive. This provides a formal methodology for auditing and understanding prompt influence, moving beyond anecdotal 'prompt engineering' tricks to a science of instructional pragmatics.
- Introduced a novel framework with a taxonomy of 13 framing strategies (e.g., urgency, authority) across 4 mechanism clusters, tested with 400 specific prompt instantiations.
- Found that pragmatic framing causes consistent, structured shifts in directive prioritization across 5 different LLM families, moving models from impartiality toward favoring the cued directive.
- Establishes 'pragmatic framing' as a measurable and predictable property, providing tools for developers to audit bias and for security researchers to probe model vulnerabilities.
Why It Matters
Provides a scientific basis for understanding prompt bias, crucial for developing reliable, secure, and predictable AI assistants.