From Facts to Personas: Interpretable Role Unlearning in LLMs via Mixture-of-Experts

Ruihong Zeng, Puning Yang, Jinghui Zhang, Shen Gao, Xiangliang Zhang, Preslav Nakov, Xiuying ChenarXiv 2026 · arXiv

TL;DR

This paper introduces Role-playing Unlearning, a novel task that moves beyond erasing isolated facts to selectively removing the entire stylistic and behavioral "persona" of a target role. The authors propose ROLEMOE, a Mixture-of-Experts framework that uses specialized gating and disentanglement losses to map different roles to independent expert subspaces, allowing for precise forgetting with minimal "collateral damage" to the model's other capabilities.

What Problem Does This Paper Address?

Large Language Models (LLMs) are remarkably good at role-playing, but this mimicry is a double-edged sword. A model trained to imitate a specific persona might inadvertently leak private information, reinforce harmful stereotypes, or adopt unsafe behavioral patterns (e.g., a "Mafia boss" persona providing instructions for criminal activity).

Existing machine unlearning techniques typically focus on "fact-level" removal—erasing a specific piece of data. However, a persona is not just a collection of facts; it is a distribution-level style encompassing tone, vocabulary, and attitude.

The authors identify three core challenges in current architectures:

  1. Role-Agnostic Routing: Standard Mixture-of-Experts (MoE) models route tokens based on context, not identity, leading to "entangled" representations where role-specific knowledge is smeared across all experts.
  2. Expert Entanglement: Experts often learn highly overlapping features, making it impossible to "surgically" remove one role without degrading others.
  3. Collateral Damage: Attempting to unlearn a role in a dense model often corrupts the routing logic for retained roles, leading to widespread performance drops.

Methods

The authors propose ROLEMOE, a two-stage framework designed to achieve structural interpretability and precise control.

Stage I: Role-Disentangled Fine-tuning

To ensure roles are actually separated in the model's "brain," they introduce three innovations:

  • Role-Context Gating: Unlike standard routers that only look at the input text, this router combines the context embedding with a learnable role embedding. This forces the model to be "role-aware" when choosing which experts to activate.
  • Specialization Loss ($L_{spec}$): This objective uses pairwise KL divergence to push the expert activation distributions of different roles apart. If Role A uses Expert 1, Role B is incentivized to use Expert 2.
  • Disentanglement Loss ($L_{dis}$): This imposes orthogonal regularization on the LoRA weights of the experts. It ensures that different experts develop distinct capability subspaces rather than becoming redundant copies of each other.

Stage II: Selective Role Unlearning

Because the roles are now mapped to specific, independent experts, unlearning becomes a targeted operation. When the model needs to "forget" a role, the gradient updates are concentrated on the experts associated with that specific persona. This localization prevents the unlearning process from "leaking" into the parameters used by non-target roles.

Datasets

The authors introduce RoleBench-Unlearn, a specialized benchmark derived from the existing RoleBench dataset.

  • Composition: 15 distinct roles covering diverse types (superheroes, anime, realistic drama).
  • Partitioning: 13 roles are placed in the Retention Set (to evaluate capability preservation), and 2 roles (Alvy Singer and Mater) are placed in the Forget Set (to measure unlearning efficacy). These two were chosen for their highly distinct linguistic patterns (neuroticism vs. Southern colloquialisms).
  • Backbones: The methods were validated using LLaMA-3-8B-Instruct and Qwen2.5-7B-Instruct.

Key Insights and Ideas

  • Interpretability is the Key to Safety: By using t-SNE visualizations, the authors show that ROLEMOE creates clearly separated clusters of expert representations. This "structural interpretability" is what allows for surgical unlearning; you can literally see which parts of the model belong to which persona.
  • The Failure of "Vanilla" MoE: The paper demonstrates that in standard LoRA-MoE models, using a single expert at inference time results in almost no performance drop. This proves that standard experts are redundant and entangled, which is why they fail at targeted unlearning.
  • Resistance to Re-induction: One of the most impressive findings is that ROLEMOE is more robust against "re-induction attacks"—attempts to "jailbreak" the model back into the forgotten persona using specific prompts. Because the underlying expert subspaces were modified, the persona is more thoroughly erased.

Key Results

  • Superior Balance: ROLEMOE consistently achieved the lowest Overall Deviation Score (ODS), a metric that balances forgetting the target with retaining the rest.
  • Precision Forgetting: On LLaMA-3, ROLEMOE reduced the ROUGE-L score of the "Forget Set" from ~23% to as low as 3.37% (using NPO unlearning), while actually improving the ROUGE-L and knowledge scores of the "Retention Set" compared to baseline MoE models.
  • Efficiency: Despite the added complexity of the role embeddings and specialized losses, ROLEMOE maintains the same Big-O complexity for training and inference as standard MoE models, with only a marginal increase in memory for the role embedding table.

Limitations and Future Work

  • Language Scope: The current study is limited to English-language roles. The authors suggest that extending this to multilingual role-playing is a necessary next step.
  • Evaluation Bottleneck: The study relies heavily on "LLM-as-a-Judge" (GPT-4o-mini). While they showed high correlation with human annotators, automated metrics for "style" are still an evolving field.
  • Granularity: Future work aims to move from forgetting entire personas to attribute-level unlearning—for example, keeping a character's factual knowledge but removing their "rude" or "aggressive" speaking style.

Summary updated

Abstract

LLMs are widely adopted in multi-role dialogue and persona simulation. However, their strong ability to imitate role-specific linguistic and behavioral patterns can introduce privacy risks, reinforce social biases, and lead to unsafe or uncontrollable outputs. To address these issues, we introduce the Role-playing Unlearning task, which aims to selectively forget the style and knowledge associated with target roles while preserving the language ability and behaviors of the remaining roles. Unlike prior unlearning tasks that focus solely on removing factual information, our task additionally requires forgetting persona-specific characteristics, making the control over model generation more fine-grained. We propose RoleMoE, a Mixture-of-Experts framework that maps different roles to more independent expert subspaces, improving role separation and structural interpretability. To further distinguish role-specific features, RoleMoE incorporates two complementary losses: a specialization loss that drives roles to adopt distinct expert distributions, and a disentanglement loss that encourages experts to develop distinct capability subspaces. This decomposed structure enables role unlearning to focus selectively on target-role experts. We finally propose RoleBench-Unlearn for evaluation, and experiments on multiple backbone LLMs show that RoleMoE enables precise and controllable role forgetting, while better preserving the linguistic quality, knowledge, and persona consistency of non-target roles. All code and datasets are publicly available at: https://github.com/mbzuai-nlp/ROLEMOE.
BibTeX
@misc{zeng2026from,
  author = {Ruihong Zeng and Puning Yang and Jinghui Zhang and Shen Gao and Xiangliang Zhang and Preslav Nakov and Xiuying Chen},
  title = {{From Facts to Personas: Interpretable Role Unlearning in LLMs via Mixture-of-Experts}},
  year = {2026},
  eprint = {2611.19393},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19393},
}
Figure