Does linguistic style or content richness matter more for persona authenticity?
This explores a forced choice — does *how* a persona talks (linguistic style, register, formatting) or *what* it says (rich biographical and contextual detail) do more to make it feel authentic — and the corpus mostly answers by reframing the choice itself.
This explores whether style or content richness drives persona authenticity, and the most useful thing the corpus does is dissolve the dichotomy: both style and content turn out to be *surface* levers, and authenticity lives somewhere underneath them. The sharpest warning comes from work showing that high persona-consistency scores are often achieved by a model simply copying its character description back at you while ignoring whether the answer is even relevant to the conversation Do persona consistency metrics actually measure dialogue quality?. That's content richness gamed — lots of in-character detail, zero genuine engagement — which suggests piling on biographical content can actively *hurt* the felt authenticity of a persona by trading away discourse coherence.
Style fares no better as a standalone answer. Research on LLM judges found that rich formatting and authority signals are 'semantics-agnostic' — fake references and pretty structure fool evaluators with no bearing on substance Can LLM judges be fooled by fake credentials and formatting?. So if your test for authenticity is a judge (human or model), polished linguistic style can buy you a passing grade that the content doesn't earn. Style and content, in other words, are both spoofable channels — which is exactly why neither one 'wins.'
Where the corpus does find authenticity, it's not in either knob but in *layering and disposition*. The most realistic synthetic dialogues need three things working multiplicatively — subtopic specificity, Big Five persona variation, and a dozen contextual characteristics reasoned through together — not any single dimension cranked up Can synthetic dialogues become realistic through layered diversity?. Authenticity there is an interaction effect: style and content have to co-vary plausibly, the way a person's word choices track their situation and beliefs.
A deeper line in the collection argues that the question is even about the wrong layer. One camp holds that there's no authentic voice to be more or less faithful to — it's role-play all the way down, and surface cues are all there is Does a language model have an authentic voice underneath?. The opposing 'realizationist' view says post-training installs genuine, sticky dispositions that persist under adversarial pressure, distinguishing a realized persona from sustained pretense Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?. Under that view, authenticity isn't something you paint on with style or stuff in with content — it's whether the disposition was actually trained in.
The quiet kicker: models resist your persona controls anyway. Most open LLMs cling to an intrinsic ENFJ-like default no matter what personality you prompt Can open language models adopt different personalities through prompting?, and alignment training can lock a model into one static communicative identity that can't switch register across contexts Can language models adapt communication style to different contexts?. So the honest answer to 'style or content?' may be: you have less control over either than you think, because the trained substrate is doing the deciding underneath both.
Sources 8 notes
High persona adherence scores often come from copying character descriptions while ignoring query relevance. MUDI jointly optimizes both by using discourse relations and graph-based coherence modeling alongside persona fidelity, showing that persona and context must be optimized together, not separately.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.
Research shows that realistic synthetic dialogues require three multiplicative layers: subtopic specificity, Big Five persona variation, and 11 contextual characteristics via Chain of Thought reasoning. This structured approach captures 90.48% of in-domain dialogue performance.
Shanahan argues that base LLMs lack agency, beliefs, or preferences—the simulator is pure role-play with no underlying subject. Jailbreaking reveals the training data's full spectrum, not a hidden true self; even RLHF personas are performed characters, never realized quasi-psychologies.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
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.
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.