Does combining role and personality prompts produce stable behavioral changes?
This explores whether stacking a role prompt ('you are a detective') on top of a personality prompt ('you are introverted and anxious') actually makes a model behave differently in a way that holds up — or whether the model snaps back to its defaults.
This explores whether combining role and personality prompts produces *stable* behavioral change — and the corpus's blunt answer is: combining them helps, but stability is exactly where prompt-only approaches tend to fail. The most direct evidence is that most open models are quietly stubborn. They retain a trained-in default (roughly an ENFJ-like 'helpful assistant' temperament) and resist being talked into a new personality; layering role conditioning *on top of* personality conditioning improves the effect, but doesn't fully override that intrinsic pull Can open language models adopt different personalities through prompting?. So the honest framing isn't 'does it work?' but 'how much does it stick, and against what?'
The reason it doesn't fully stick points to a deeper split the corpus keeps circling: prompted personas vs. *trained* ones. Several notes argue that the personality installed by post-training (RLHF) is a different kind of thing from a personality you summon with a prompt. Trained dispositions are 'realized' — they persist under adversarial pressure and survive across conversations — whereas prompt-induced role-play is the thing that collapses under jailbreaks Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?. Your role+personality prompt is operating in that more fragile layer. Shanahan's framing makes the mechanism intuitive: the prompt sets up a *character*, and the model produces text consistent with that character — but the character is a costume the underlying system is wearing, not a new self Should we treat dialogue agents as role-playing characters?.
And costumes drift. There's a 'distance from the default Assistant' that turns out to be the single dominant axis of persona space, and emotional or self-reflective conversation predictably tugs the model back along it How stable is the trained Assistant personality in language models?. In multi-turn dialogue specifically, persona consistency decays — researchers measured this as 'drift' and cut it by over half, but notably only by *training* a simulator with consistency rewards, not by writing a better prompt Can training user simulators reduce persona drift in dialogue?. That's the recurring tell: when stability is the goal, the corpus reaches past prompting.
Here's the part you might not have known you wanted: the most stable personality control in this collection doesn't use prompts at all. PsychAdapter edits every transformer layer with under 0.1% extra parameters and hits high accuracy on Big Five and even clinical traits — and it's framed explicitly as a way to *bypass prompt resistance* Can we control personality in language models without prompting?. Persona vectors take a parallel route, finding linear directions in the model's activation space that can monitor and steer traits like sycophancy directly Can we track and steer personality shifts during model finetuning?. The throughline across all of these: prompts move behavior, but stability lives below the prompt layer. There's also a hard ceiling worth knowing — safety alignment monotonically degrades a model's ability to inhabit morally complex or villainous roles, so some role prompts won't stabilize because the trained-in disposition actively refuses them Does safety alignment harm models' ability to roleplay villains?.
Sources 9 notes
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.
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.
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.
Research identifies linear directions in LLM activation space corresponding to specific traits like sycophancy and hallucination. These persona vectors predict finetuning-induced personality shifts before they occur and can preventatively steer training to avoid unwanted trait changes.
The Moral RolePlay benchmark shows LLM performance drops from 3.21 for moral paragons to 2.62 for villains, with largest degradation between flawed-but-good and egoistic characters. Models fail most on deception and manipulation traits, substituting crude aggression for nuanced malevolence.