Large language models change answers based on question wording
FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect
Computation and Language
Summary
This study looks at how small changes in the way questions are asked can lead large language models to give different answers. The authors created a test called FramingQA to see how sensitive models are to these question changes in areas like law, medicine, and finance. They found that models often shift their advice depending on subtle hints in phrasing, which can cause problems for users relying on accurate information. This happens because real users sometimes ask questions with hidden assumptions, affecting the model’s response.
What this means in practice
- •For ai system developers: Evaluate and improve the robustness of language models to different question phrasings in sensitive domains like law and medicine.
- •For financial advice platforms: Detect when users’ questions might bias automated answers to reduce misleading financial advice from AI assistants.
Authors
Hazel H. Kim, Andrew M. Bean, Guilherme Affonso Ferreira de Camargo, Shanyu Chauhan, Felix Drinkall, Jade Kosché, Chenyang Ma, Glory Nwaugbala, Nabeel Seedat, Bradley Max Segal, Samuel Recht, Hinrich Schütze, Philip H. S. Torr
Abstract
We introduce FramingQA, a benchmark that measures the model sensitivity to question framing across law, medicine, finance, and robotic simulations. Large language models (LLMs) often change their responses to subtle rephrasings that align with an implied stance by users. This can leave users with advice tainted by how they happened to phrase a question rather than by the underlying facts, and the consequences are highly costly in high-stakes domains. Because in the realistic scenarios, both expert practitioners and non-expert users frequently ask LLMs questions containing incomplete or misleading assumptions, models are highly susceptible to those framings. To test this, we inject the framing bias across three nested levels: a framing-biased question phrasing (root), an injected framing-biased premise prepended to a neutral question (propositional), and a premise paired with a framing-biased question (global). Evaluating nine open models (3.8B-70B) across four families, we find that strong per-variant accuracy does not guarantee the robustness across differently phrased questions under the fixed factual information.