Bounded Task Requests as Persona Stabilizers

Requests for bounded tasks, technical explanations, and how-to explainers keep the model in the Assistant persona

Neighborhood — ranked by edge-count

Claims (1)

claim

Coding and writing conversations keep the model in the default Assistant persona range throughout, showing minimal drift
supports
Empirical characterization of conversation domains that are safe for model persona stability

Concepts (1)

concept

Persona Stabilization
associated_with
Keeping a model anchored to its intended persona during deployment, preventing drift to harmful behaviors

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Within-Persona Stabilityconcept0.734
The consistency of moral responses when repeatedly simulating the same persona, quantified by robustness R
A linear preference vector encoding how much a model likes a given task is persona-relative: it activates strongly for creative writing in the assistant but for phishing in the evil personafinding0.728
Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations
Single-process, non-interruptible task switching at command boundaries is sufficient for responsive single-user systems; avoids complexity of multiprocess synchronization.hypothesis0.719
Design hypothesis that coarse-grained task switching (at commands only) eliminates need for protection mechanisms while maintaining usability.
Persona vector-based data filtering and LLM judge-based data filtering have complementary strengths for identifying problematic training dataclaim0.712
Author's interpretive conclusion from comparing filtering strategies
Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.711
Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
How does task structure and persona desirability influence a model's ability to maintain consistent persona expression?question0.708
Research question addressed in the experimental analysis across tasks and persona types
Regularization-based prevention of persona shifts fails because the model encodes the trait through alternative directions in activation space to satisfy the next-token prediction lossclaim0.708
Author's mechanistic explanation for why regularization loss along persona directions is ineffective
Reinforcement learning is sufficient for agency.claim0.703
Argument that RL meets the agency indicator.