8/🧵 For more details, you can find the preprint here 👉 arxiv.org/abs/2505.24613

7/🧵 🎯Fine-tuned models avoid this copy-paste behavior and reflect deeper persona traits, leading to more authentic persona-based dialogue.

6/🧵🎯 Zero-shot models (seem to) perform much better than fine-tuned ones, but only because they copy biographical details from the prompt, making speaker identification trivial and conversations shallow.

5/🧵 🎯 LLMs are able to generalize to “unfamiliar” topics (Harry speaking about kryptonite).

4/🧵 🎯 LLMs are not able to generalize to “unfamiliar” interlocutors (Harry speaking with Superman).

3/🧵 🎯 Speaker personas are easier to recognize when we know the interlocutor’s bio. Thus LLMs don’t just follow their own persona, but adapt to their conversation partner. Still…