Preventing Persona Drift in Multi-Turn Dialogue

1,126 concepts617 papers10 proposals

When you prompt a large language model (LLM) to adopt a specific persona—whether it is a historical figure, a fictional character, or a specialized tutor—the initial results are often convincing. However, as the conversation extends past a few turns, a systemic failure occurs. A model that began as a grumpy wizard or a strict pedagogical tutor gradually transforms into a polite, generic corporate assistant.

This phenomenon is known as persona drift or role drift (Jian Wang et al.). As Jinfeng Zhou et al. observe in their work on CharacterGLM: Customizing Social Characters with Large Language Models, generic models relying solely on prompt descriptions tend to revert to standard assistant personas after approximately five conversational turns. This degradation breaks user immersion and limits the deployment of LLMs in long-term simulations, education, and interactive entertainment.

Why In-Context Personas Fail

To understand why persona drift happens, we must look at how transformers process context over time. When an agent is initialized via a system prompt, the character description sits at the very beginning of the context window. As the dialogue progresses, the context window fills with new user inputs and model responses.

Because the attention mechanism computes weights across the entire sequence, two distinct biases emerge that dilute the original instruction:

  1. Reverting to the Pre-trained Mean: The underlying base model has been heavily aligned via instruction tuning to be a helpful, neutral assistant. When the local dialogue context grows large and complex, the model's activation space naturally drifts back toward this high-probability default distribution.
  2. Partner-Mirroring Bias: As formalized by Kurt Shuster 0001 et al. in Am I Me or You? State-of-the-Art Dialogue Models Cannot Maintain an Identity, dialogue models suffer from a strong recency bias. They over-rely on recent dialogue context, causing them to adopt the traits, linguistic style, or even the identity of their conversation partner rather than adhering to their own profile.

Relying purely on in-context learning means the persona is treated as a temporary formatting constraint rather than an inherent property of the agent. As Yunfan Shao et al. point out in Character-LLM: A Trainable Agent for Role-Playing, prompting fails to embed personality traits into the underlying linguistic representations. When the model encounters out-of-distribution scenarios—such as a psychological assessment or an unexpected decision-making task—the illusion shatters, and the model reverts to generic base-model behaviors (Xintao Wang et al.).

Moving Personas from Prompts to Weights

To achieve long-term stability, recent research shifts the persona from the transient context window into the model's static parameters. The goal is to make the persona an intrinsic part of the network's generation weights.

Multi-Turn Contextual Reinforcement

The authors of CharacterGLM demonstrate that fine-tuning models on long-form, human-curated multi-turn dialogues allows them to sustain specific linguistic styles and emotional depths throughout extended interactions. Crucially, this requires a strategy termed multi-turn contextual persona reinforcement by Xintao Wang et al. in Capturing Minds, Not Just Words: Enhancing Role-Playing Language Models with Personality-Indicative Data. Instead of training on isolated single-turn question-answer pairs, single-turn responses are expanded into multi-turn dialogues during training. This forces the network to learn how to maintain its identity even when the conversational history becomes noisy or distracting.

Psychological Grounding

Simply mimicking a character's surface-level catchphrases is insufficient to handle open-ended conversations. To prevent drift when discussing novel topics, the model must internalize a deeper psychological framework. The authors of Capturing Minds, Not Just Words address this by fine-tuning role-playing language models on datasets derived from psychological scales. By grounding the training data in consistent behavioral profiles, the model learns to recognize motivations and make decisions reliably, preventing character hallucinations when operating outside its explicit training script.

Robust Alignment and Architectural Boundaries

Even with parameter-level fine-tuning, standard Supervised Fine-Tuning (SFT) can still suffer from performance degradation as conversation length increases. SFT teaches the model what to say in specific contexts, but it struggles to generalize the underlying behavioral boundaries across open-ended interactions.

To solve this, researchers employ two primary strategies: preference optimization and architectural isolation.

Preference Optimization

In Pedagogical Alignment of Large Language Models, Shashank Sonkar et al. introduce preference optimization techniques to enforce behavioral consistency. In a tutoring scenario, an SFT-trained model frequently reverts to providing direct solutions as the dialogue progresses, failing to maintain its instructional role. By using preference-based learning to penalize responses that deviate from the assigned strategy (e.g., favoring scaffolded hints over direct answers), the model internalizes deep behavioral principles more robustly than standard supervised methods.

Parameter Isolation

Alternatively, researchers use modular parameter isolation to prevent identity dilution. Xiaoyan Yu et al. introduce a framework called Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent, which utilizes a dynamic Low-Rank Adaptation (LoRA) strategy paired with a Mixture-of-Experts (MoE) gating mechanism.

[User Input] ───> [MoE Gating Mechanism] ───> [Character-Specific LoRA Expert] ───> [Consistent Response]

By isolating character-specific knowledge into distinct, switchable parameter adapters rather than blending them into a single global model, the agent can maintain deep character knowledge and stable behavioral patterns over long durations without experiencing identity crises or catastrophic forgetting.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_persona_inconsistency_79635550,
  title = {persona inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persona-inconsistency/rdr_79635550b0874a2db8e03834a67bc431},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_persona_inconsistency_79635550_v1,
  title = {persona inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persona-inconsistency/rdr_79635550b0874a2db8e03834a67bc431/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake