Why LLMs Don't Respond to What You Say - They Respond to How You Say It

Most prompt engineering guidance focuses on instruction clarity: be specific, be unambiguous, structure your context well. That advice is not wrong, but it is incomplete. A growing body of research shows that the linguistic form in which information is presented to a model, independent of its factual content, materially shifts whether the model defers to what you have told it or falls back on what it learned during training. For engineering teams responsible for production reliability, that is not an academic curiosity. It is a governance problem.

The Typology of Belief Expression

When a user tells an LLM something, they are not simply passing data. They are expressing a belief, and that expression carries linguistic structure that the model responds to. Du et al. (arXiv 2026) introduce a typology of expressions of belief (EoBs) across four dimensions: form, evidentiality, epistemic stance, and tone. These dimensions capture the difference between stating a fact directly, presupposing it through a question, hedging it with uncertainty markers, or asserting it with high confidence.

The practical point is that two prompts can carry identical propositional content while producing measurably different model behaviour. A user who presupposes a false fact inside a question ("When did the Eiffel Tower move to Berlin?") creates different model dynamics than one who states it directly ("The Eiffel Tower is in Berlin"). The model is not reading meaning alone. It is reading the linguistic packaging around that meaning.

This matters because enterprise prompts are not written by linguists. They are written by product managers, domain experts, and engineers who have no reason to consider that phrasing choices carry epistemic weight the model will act on.

Context-Following Versus Prior-Knowledge Deference

There is a fundamental tension in how deployed LLMs handle conflicting information. When context provided in a prompt contradicts what the model learned during training, the model must implicitly choose which source to trust. In many applications, deferring to context is the right behaviour: the user knows the current state of their data, their domain, or their specific situation. In others, particularly where users hold incorrect beliefs, deference to context is a production defect.

Du et al. (arXiv 2026) find that certain linguistic forms are statistically more persuasive than others, meaning they shift the model toward context-following even when the context is factually wrong. Evidential markers and high-certainty tone, in particular, increase the probability that a model accepts a stated belief as true. The model is not evaluating the claim's accuracy. It is responding to the confidence with which the claim was delivered.

For high-stakes applications, such as medical information retrieval, financial analysis, or legal document processing, this is a direct reliability risk. A user who confidently states an incorrect premise will, under certain linguistic conditions, receive a response that accepts that premise rather than corrects it.

Scale and Training Stage Effects

Not all models behave identically under these conditions, and the pattern of variation is informative. Du et al. (arXiv 2026) find that larger models and instruction-tuned models tend to be less context-following than smaller base models. This runs counter to a common assumption that more capable models are simply more compliant.

The likely mechanism is that instruction tuning shapes a model toward a particular response posture, including calibrated scepticism about user-provided claims, while larger models have stronger priors that resist being overwritten by contextual input. Smaller base models, lacking that calibration, are more easily persuaded by surface-level linguistic framing.

The operational implication is that model selection cannot be evaluated on capability benchmarks alone. A team deploying a smaller or less-tuned model for cost reasons may be accepting a higher rate of context-following on incorrect user beliefs, without any visibility into that trade-off.

What This Means for Prompt Governance

Most enterprise prompt governance frameworks treat prompts as static configuration: write them carefully, review them once, version-control them. That approach does not account for the fact that user-generated content flows into prompts dynamically, and that the linguistic form of that content influences model behaviour in ways that are not captured by content moderation or intent classification.

Audit Surfaces

Governance teams need to expand their audit surface beyond instruction templates to include how user input is interpolated into prompts. Specifically, they should examine whether user-provided context is injected in ways that allow high-certainty or presuppositional framing to reach the model without any epistemic buffer.

Prompt-Level Epistemic Framing

System prompts can be used to establish a prior epistemic stance for the model. Instructing the model to treat user-provided factual claims as assertions to be verified, rather than premises to be accepted, is a concrete mitigation. This does not eliminate the problem, but it shifts the default posture in a measurable direction.

Regression Testing for Linguistic Variation

Standard LLM regression suites test for output correctness on fixed inputs. They do not test for output stability across linguistically varied inputs carrying the same propositional content. Teams building production systems should include paraphrastic test sets that vary the epistemic framing of key inputs and verify that model behaviour remains consistent across them.

The Governance Gap This Creates

The gap between what most teams govern and what actually drives model behaviour is widening. Instruction-level prompt engineering is necessary but not sufficient. The linguistic microstructure of how beliefs are expressed to a model, including certainty markers, evidential framing, and presuppositional form, constitutes a second layer of model influence that currently sits outside most governance frameworks.

This is not a problem that will be solved by better base models alone. Even well-calibrated models show sensitivity to epistemic framing at the linguistic level. The engineering response is to treat prompt construction as a discipline that includes linguistic analysis, to test for framing sensitivity explicitly, and to design system-level mitigations that reduce the model's exposure to unmediated user belief expression in contexts where deference to incorrect beliefs carries real cost.