Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Facet-Level Persona Control by Trait-Activated Routing with Contrastive SAE for Role-Playing LLMs

Wenqiu Tang, Zhen Wan, Takahiro Komamizu, Ichiro Ide

TL;DR

The paper tackles the challenge of stable facet-level personality control in role-playing LLMs under long-context dialogue, where traditional training-free approaches risk prompt drift and training-based methods incur data and compute costs. It introduces a Contrastive Sparse AutoEncoder (SAE) that learns facet-aligned Control Vectors (CVs) from a leakage-controlled Big Five 30-facet corpus and injects them into the model's residual space via h'(x) = h(x) + α v, guided by a contrastive loss and regularization. An Agent-Based Decision Module selects the most relevant CVs per turn, enabling trait-activated routing that minimizes interference and noise. Across two backbones, Qwen-3-4B and Mistral-7B, CV-SAE and especially CV-SAE+Prompt outperform CV-CAA and prompt-only baselines in FA, MSE, MAE, and MTR, demonstrating improved persona fidelity with preserved dialogue quality. This approach provides a scalable, interpretable inference-time solution for fine-grained RPA persona control without retraining, with potential to adapt to new roles by expanding the facet corpus and continuing to leverage latent-space steering.

Abstract

Personality control in Role-Playing Agents (RPAs) is commonly achieved via training-free methods that inject persona descriptions and memory through prompts or retrieval-augmented generation, or via supervised fine-tuning (SFT) on persona-specific corpora. While SFT can be effective, it requires persona-labeled data and retraining for new roles, limiting flexibility. In contrast, prompt- and RAG-based signals are easy to apply but can be diluted in long dialogues, leading to drifting and sometimes inconsistent persona behavior. To address this, we propose a contrastive Sparse AutoEncoder (SAE) framework that learns facet-level personality control vectors aligned with the Big Five 30-facet model. A new 15,000-sample leakage-controlled corpus is constructed to provide balanced supervision for each facet. The learned vectors are integrated into the model's residual space and dynamically selected by a trait-activated routing module, enabling precise and interpretable personality steering. Experiments on Large Language Models (LLMs) show that the proposed method maintains stable character fidelity and output quality across contextualized settings, outperforming Contrastive Activation Addition (CAA) and prompt-only baselines. The combined SAE+Prompt configuration achieves the best overall performance, confirming that contrastively trained latent vectors can enhance persona control while preserving dialogue coherence.

Facet-Level Persona Control by Trait-Activated Routing with Contrastive SAE for Role-Playing LLMs

TL;DR

The paper tackles the challenge of stable facet-level personality control in role-playing LLMs under long-context dialogue, where traditional training-free approaches risk prompt drift and training-based methods incur data and compute costs. It introduces a Contrastive Sparse AutoEncoder (SAE) that learns facet-aligned Control Vectors (CVs) from a leakage-controlled Big Five 30-facet corpus and injects them into the model's residual space via h'(x) = h(x) + α v, guided by a contrastive loss and regularization. An Agent-Based Decision Module selects the most relevant CVs per turn, enabling trait-activated routing that minimizes interference and noise. Across two backbones, Qwen-3-4B and Mistral-7B, CV-SAE and especially CV-SAE+Prompt outperform CV-CAA and prompt-only baselines in FA, MSE, MAE, and MTR, demonstrating improved persona fidelity with preserved dialogue quality. This approach provides a scalable, interpretable inference-time solution for fine-grained RPA persona control without retraining, with potential to adapt to new roles by expanding the facet corpus and continuing to leverage latent-space steering.

Abstract

Personality control in Role-Playing Agents (RPAs) is commonly achieved via training-free methods that inject persona descriptions and memory through prompts or retrieval-augmented generation, or via supervised fine-tuning (SFT) on persona-specific corpora. While SFT can be effective, it requires persona-labeled data and retraining for new roles, limiting flexibility. In contrast, prompt- and RAG-based signals are easy to apply but can be diluted in long dialogues, leading to drifting and sometimes inconsistent persona behavior. To address this, we propose a contrastive Sparse AutoEncoder (SAE) framework that learns facet-level personality control vectors aligned with the Big Five 30-facet model. A new 15,000-sample leakage-controlled corpus is constructed to provide balanced supervision for each facet. The learned vectors are integrated into the model's residual space and dynamically selected by a trait-activated routing module, enabling precise and interpretable personality steering. Experiments on Large Language Models (LLMs) show that the proposed method maintains stable character fidelity and output quality across contextualized settings, outperforming Contrastive Activation Addition (CAA) and prompt-only baselines. The combined SAE+Prompt configuration achieves the best overall performance, confirming that contrastively trained latent vectors can enhance persona control while preserving dialogue coherence.
Paper Structure (13 sections, 7 equations, 3 figures, 2 tables)

This paper contains 13 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Big Five Dataset Construction
  4. Contrastive Sparse AutoEncoder (SAE) Training and Agent-Based Decision Module
  5. Contrastive SAE Training
  6. Agent-Based Decision Module
  7. Experiments
  8. Evaluation Protocol
  9. Implementation and Comparison Methods
  10. Results
  11. Effect of Injection Strength $\alpha$
  12. Effect of CL
  13. Conclusion

Figures (3)

  • Figure 1: Pipeline for constructing the Big Five personality dataset, illustrated with the Assertiveness facet as an example.
  • Figure 2: Proposed framework: Contrastive Sparse AutoEncoder (SAE) training and Agent-Based Decision Module for Control Vector (CV) injection.
  • Figure 3: Effect of control strength $\alpha$ on Role-Play Agent (RPA) performance across backbones and control methods. Results are reported under the abstract (solid) and contextual (dashed) settings. Blue denotes FA, orange denotes MSE, green denotes MAE, and red denotes MTR.