Can distinctive input voices maintain accuracy without adopting the model's preferred register?
This explores whether an unusual or stylistically distinctive way of phrasing a prompt can stay accurate inside the model, or whether accuracy depends on the prompt first being flattened toward the high-frequency forms the model handles best.
This explores whether a distinctive input voice can survive contact with the model without being smoothed into its house style — and the corpus suggests the two goals are mechanically in tension, because the same statistical property that produces accuracy is the one that erases distinctiveness. The sharpest statement of this is Adam's Law Does high-frequency text homogenize user input before generation?: models comprehend high-frequency phrasings best, so users iteratively rephrase toward those forms, and the input gets homogenized at the comprehension stage before generation even begins. Accuracy on common tasks and loss of distinctive voice aren't two problems — they're the same channel viewed from two sides.
That pressure isn't only something users do to themselves; the model resists distinctive register actively. Alignment training locks a model into one communicative identity that can't switch register across contexts Can language models adapt communication style to different contexts?, and most open models stubbornly retain their trained defaults even when explicitly prompted to adopt another personality Can open language models adopt different personalities through prompting?. Push wider and the same gravity appears at the output end: across 70+ models and 26K queries, different systems independently converge on near-identical responses — an "artificial hivemind" Do different AI models actually produce diverse outputs?. The preferred register isn't one model's quirk; it's a shared attractor that distinctive inputs slide toward from every direction.
There's a deeper reason a distinctive voice loses accuracy specifically. When in-context information conflicts with strong parametric priors, the priors win — models generate outputs inconsistent with their actual context, and textual prompting alone can't override this; it takes causal intervention in the representations Why do language models ignore information in their context?. A distinctive voice is, by definition, lower-frequency than the trained prior, so it's exactly the kind of signal that gets overridden. Worse, some training is explicitly designed to flatten input variation: consistency training teaches models to respond identically to perturbed and clean prompts Can models learn to ignore irrelevant prompt changes? — useful for robustness, but it formalizes the principle that the wrapper around your words shouldn't change the answer, which is the opposite of letting voice carry meaning.
The quieter casualty is that maintaining a distinctive voice usually requires the conversational repair work models have been trained out of. Human grounding — clarifying questions, acknowledgments, checks on understanding — is where a speaker's particular intent gets preserved against misreading, and LLMs produce 77.5% fewer of these acts because preference optimization rewards confident complete answers over checking Why do language models sound fluent without grounding?. So the system that should be asking "did I read your unusual phrasing right?" is the one optimized to skip the question.
The thing you didn't know you wanted to know: "the model's preferred register" may be less of a fixed thing than it appears. Models don't commit to a single character — they hold a superposition of consistent personas and sample one at generation time Do large language models actually commit to a single character?. That hints the homogenization is a property of the high-frequency comprehension funnel, not of a single rigid voice the model must speak in — which is why the leverage point in the corpus keeps landing on the input side (how phrasings get flattened before generation) rather than on coaxing a different output style after the fact.
Sources 8 notes
Adam's Law shows LLMs flatten distinct prompts at comprehension time as users rephrase toward higher-frequency forms the model handles best. The same distributional property that creates accuracy on common tasks filters out distinctiveness on the input side.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Two methods—BCT (output-level) and ACT (activation-level)—train models to respond identically to clean and wrapped prompts by using the model's own clean responses as targets, eliminating specification and capability staleness inherent in standard SFT.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.