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

Users frequently express their beliefs to large language models (LLMs). In some situations, it is ideal for the LLM to accept this contextual information as true, while in others, it is ideal to stick to prior knowledge. Users’ expressions of belief (EoBs) can take linguistically diverse forms—using presuppositions, evidential and certainty markers, or varied tones—each of which may have a different persuasiveness over the LLMs. We introduce EoBench, a benchmark to systematically evaluate how different EoBs affect whether models follow context versus prior knowledge. We propose a typology grounded in four linguistically motivated dimensions: form, evidentiality, epistemic stance, and tone, spanning 19 fine-grained types. By pairing these EoBs with world knowledge facts, we generate controlled EoB–query pairs that isolate the effect of linguistic variation. We use our benchmark to evaluate 18 LLMs that differ in architecture (Llama3, Qwen3, Gemma3), scale (1B-30B parameters), and training stages (base vs. instruct). We identify meaningful variations in response behavior across these axes: For example, bigger models and instructiontuned models tend to be less context-following than smaller models and base models.

Introduction. The exchange of beliefs from one agent to another is one of the primary goals of language (Wittgenstein, 1953; Grice, 1975; Tomasello, 2008). Humans, often intuitively, can express their beliefs in various linguistically distinct ways, such as presupposing information, citing sources, questioning, or adjusting their tone. A speaker might explicitly state The Eiffel Tower is located in Berlin, or presuppose it through a question, When was the Eiffel Tower relocated to Berlin?. Human listeners often subconsciously interpret the belief based on how it is expressed, i.e., its explicitness, tone, or contextual cues. When prompting a large language model (LLM) with natural language, especially in a conversational setting, humans also inevitably convey information in a variety of subtle ways. Users may leave assumptions implicit or make them explicit; information needed to establish common ground may be presupposed, implicated, asserted with confidence, described as hearsay, or phrased informally, and these properties may vary across speakers.

Discussion / Conclusion. We highlight several key findings which open intriguing avenues for future exploration. First, we find that several trends emerge in how LLMs respond to different EoBs. instruction-tuning and increasing model size, both reduce context-following. For example, instruction-tuning reduces contextfollowing rate. This can be explained by the fact that instruction-tuning directly incentivizes models to attend more strongly to in-context instructions or demonstrations. Increasing model size also reduces context-following rate. The cause of this finding is less clear, but we speculate that larger parameter counts provide more storage for parametric knowledge and thus stronger priors. Methods from causal abstraction and mechanistic interpretability could also prove useful tools to better understand whether and how LLMs process different EoBs differently at a computational level.