What determines the persona plasticity (α) in real agents? Is it just dependent on model's training (like how sycophantic model is?), or the agent's system prompt, or how much interaction history during the run stays in context, or the network the agent is in? Did your study from GPT-4o and Haiku 4.5 in naming games showed different α? If yes, then what do you think caused the difference of plasticity in both of the models?
What determines the persona plasticity (α) in real agents? Is it just dependent on model's training (like how sycophantic model is?), or the agent's system prompt, or how much interaction history during the run stays in context, or the network the agent is in? Did your study from GPT-4o and Haiku 4.5 in naming games showed different α? If yes, then what do you think caused the difference of plasticity in both of the models?