Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions

TL;DR

This paper presents a new evaluation protocol for long, multi-round dialogues with persona-based LLMs, revealing that persona fidelity diminishes over time and highlighting challenges in maintaining safety and instruction adherence.

Contribution

It introduces a robust long-dialogue evaluation framework and systematically studies how dialogue length impacts persona fidelity, instruction following, and safety in state-of-the-art LLMs.

Findings

Persona fidelity degrades over extended dialogues.

A trade-off exists between persona fidelity and instruction following.

Baseline models sometimes outperform persona-assigned models in early dialogue stages.

Abstract

Persona-assigned large language models (LLMs) are used in domains such as education, healthcare, and sociodemographic simulation. Yet, they are typically evaluated only in short, single-round settings that do not reflect real-world usage. We introduce an evaluation protocol that combines long persona dialogues (over 100 rounds) and evaluation datasets to create dialogue-conditioned benchmarks that can robustly measure long-context effects. We then investigate the effects of dialogue length on persona fidelity, instruction-following, and safety of seven state-of-the-art open- and closed-weight LLMs. We find that persona fidelity degrades over the course of dialogues, especially in goal-oriented conversations, where models must sustain both persona fidelity and instruction following. We identify a trade-off between fidelity and instruction following, with non-persona baselines initially outperforming persona-assigned models; as dialogues progress and fidelity fades, persona responses become increasingly similar to baseline responses. Our findings highlight the fragility of persona applications in extended interactions and our work provides a protocol to systematically measure such failures.