A new study introduces EoBench—a benchmark of ~66k false statements across 19 stylistic variations designed to test how LLMs respond to misleading information. Tested on 18 models from the Gemma, Llama, and Qwen families, the researchers found that the same falsehood affects models differently depending on tone, confidence, and grammatical form. Commands, formal language, baby talk, authority references, and confident delivery were the most persuasive styles, while weak formulations and counterfactual statements performed worst. Larger models showed more resistance to false context, and instruction tuning generally improved robustness. The authors conclude that LLM evaluation must account for query phrasing—small changes in wording can measurably shift model responses.

It’s Not What You Say, It’s How You Say It: Evaluating LLM Responses to Expressions of Belief