It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief
The paper tests how phrasing alone can make LLMs accept a user’s stated belief over their own prior knowledge.
The authors build a typology of belief expressions across form, evidentiality, epistemic stance, and tone. They use controlled belief-query pairs to isolate linguistic effects, then evaluate 16 models across architecture, size, and training stage. Larger and instruction-tuned models were generally less likely to follow the user’s context than smaller and base models. Some belief framings were statistically more persuasive than others. ArXiv · AI/CL/LG's note
The authors build a typology of belief expressions across form, evidentiality, epistemic stance, and tone. They use controlled belief-query pairs to isolate linguistic effects, then evaluate 16 models across architecture, size, and training stage. Larger and instruction-tuned models were generally less likely to follow the user’s context than smaller and base models. Some belief framings were statistically more persuasive than others. ArXiv · AI/CL/LG's note
score 4