Persona-Pruner: Sculpting Lightweight Models for Role-Playing

Abstract: LLMs (LMs) have shown remarkable potential as role-playing chatbots, delivering consistent, stylized interactions when given a specification of a character or user persona. However, applying these capabilities to real-world applications (e.g., ecosystems with numerous NPCs interacting simultaneously) exposes a critical inefficiency due to the excessive computational cost. In this paper, we question the necessity of dedicating a full, generalist model to a single persona, hypothesizing that a specific character identity relies on only a fraction of the model's total capacity. We observe that naively pruning LMs often severely degrades the role-playing performance for a specific persona; it does not distinguish between redundant knowledge and essential character traits. We propose Persona-Pruner, a framework that sculpts a lightweight role-playing model by isolating persona-specific sub-networks from a single description. Our experiments consistently show that Persona-Pruner preserves role-playing performance substantially more effectively than existing state-of-the-art LLM pruning techniques, reducing the performance drop from the dense model by up to 93.8% over the strongest baseline on RoleBench in LLM-as-a-judge score, while still maintaining general LLM capabilities. Code is available at https://github.com/jsu-kim/Persona-Pruner.

• The paper introduces a novel pruning method that extracts persona-aligned sub-networks using structured binary masks.
• It leverages automated persona-driven data synthesis to create calibration sets, achieving low degradation even at high sparsity levels.
• Experiments demonstrate that Persona-Pruner maintains role-playing fidelity and general LLM utility, enabling cost-efficient multi-agent deployments.

Persona-Pruner: Sculpting Lightweight Role-Playing LLMs

Introduction

Persona-Pruner introduces a targeted model pruning framework for LLMs, enabling efficient instantiation of lightweight, high-fidelity role-playing agents from a single foundational model. The fundamental hypothesis is that persona-specific behavior is encoded within a localized sub-network of an LLM, not its full parameter set. To realize this, Persona-Pruner couples persona-driven data synthesis with differentiable, structured pruning, extracting a persona-aligned sub-network capable of maintaining both character fidelity and general LM utility. This approach directly addresses critical constraints in multi-persona deployments—particularly for agent-rich environments like games or personalized assistants—by dramatically reducing compute and memory costs compared to model duplication or fine-tuning per persona.

Figure 1: Overview of Persona-Pruner, integrating persona-aligned data synthesis with persona-specific structured pruning to extract sub-networks within a general LLM.

Persona-Driven Data Synthesis

A central innovation is the fully automated synthesis of persona-aligned instruction datasets, in the absence of role-specific dialogue corpora. For a textual persona definition $P_{\text{target}}$ and a generic instruction dataset $D_{\text{base}}$, the pipeline first filters “persona-sensitive” queries—those that induce divergent internal representations for different personas, quantified via cross-persona representation divergence over transformer blocks. Top-scoring queries are paired with persona-consistent responses, constructed by rewriting existing generic answers under strict semantic preservation, guided by high-capacity LLMs (e.g., Llama-3.1-70B-Instruct). The result is a high-signal calibration set $D_{\text{syn}}$ suitable for mask optimization.

This approach decouples persona specialization from the availability of bespoke corpora, enabling persona instantiation from arbitrary descriptions. Furthermore, tailored datasets such as Alpaca-P, constructed by persona-driven transformation of Alpaca [alpaca], underpin rigorous benchmark evaluation.

Persona Sub-network Discovery

Given $D_{\text{syn}}$, Persona-Pruner performs persona-specific sub-network discovery by learning structured binary masks over the FFN intermediate dimensions within each transformer block. The mask optimization objective minimizes the negative log-likelihood of generating persona-stylized responses, with model weights frozen and only mask parameters updated. Mask binarization employs the Straight-Through Estimator (STE) for tractable gradient flow, enforcing target sparsity constraints.

This structured pruning targets the highest-parameter components—FFN intermediates—yielding sub-networks that are directly deployable on standard hardware. Mask optimization for each persona is independent, allowing efficient parallel agent instantiation.

Experimental Results

Quantitative Role-Play & Generalization

Empirical results demonstrate clear superiority over prior pruning baselines (Depth Pruning, SliceGPT, LLM-Pruner, Adapt-Pruner, etc.), both without and with post-pruning recovery finetuning. At 25% sparsity, Persona-Pruner exhibits negligible degradation in LLM-as-a-judge persona scores (≤3% drop versus dense), and at 50% sparsity, it outperforms all baselines by wide margins, even retaining competitive general LLM capabilities (OBQA/PIQA accuracy on par with dense). Notably, under recovery finetuning, Persona-Pruner often matches or exceeds the parent model’s role-playing performance.

Figure 2: Data scalability analysis demonstrating monotonic improvements in role-playing performance with increasing synthesis set size and data diversity.

Mask Stability and Semantic Alignment

Extensive analysis reveals that learned pruning masks are both reproducible across runs (cross-seed Jaccard similarity $r=0.99$) and semantically aligned (pairwise mask similarity correlates with persona embedding similarity, $r=0.65$), supporting the claim that LLMs organize persona behavioral knowledge into locally distinct sub-networks.

Figure 3: Cross-seed alignment—mask similarity for 10 personas across multiple seeds; high reproducibility indicates mask structure stability.

Figure 4: Pairwise Jaccard similarity heatmaps for persona-specific binary pruning masks across multiple seeds, demonstrating distinct, reproducible sub-network structure per persona.

Qualitative and Human Alignment Analysis

Qualitative evaluation consistently shows Persona-Pruner generating persona responses with deeply integrated character traits—beyond superficial attribute mentions, adjusting tone and content per description—contrasting with baselines that devolve into generic or attribute-hallucinated responses after pruning. Human evaluation (CoSER) places Persona-Pruner at top rank across anthropomorphism, character fidelity, and storyline quality, in strong agreement with LLM-based automated judging (Spearman $\rho=0.87$).

Multi-turn Consistency

Unlike baseline-pruned models, Persona-Pruner maintains persona attribute consistency across multi-turn dialogues, reliably preserving style and characteristic responses over long interaction spans.

Figure 5: Multi-turn persona consistency over 10 dialogue turns, with Persona-Pruner maintaining higher attribute stability than baseline-pruned models.

Theoretical Implications

These results provide strong empirical evidence for functional compartmentalization of persona-relevant computation in LLMs—supporting the notion that identity-conditioned behaviors can be mapped to sparse, disentangled subspaces within model parameterization. From a network interpretability perspective, this straightforwardly enables analyses of role-dependent capacity allocation and offers mechanisms for persona composition, modulation, or transfer in future work.

Practical Implications and Future Directions

Persona-Pruner unlocks significant cost and scaling advantages for real-world, multi-agent applications wherein hundreds or thousands of concurrent role-playing agents are required. The persona-driven data synthesis approach circumvents expensive data collection or manual annotation, supporting scalable, automated agent deployment.

For future research, several avenues are immediately promising:

• Dynamic persona mixtures: Composing or interpolating between masks for blended or transient personas.
• Adaptive inference: Runtime persona switching or mixture-of-mask selection for contextually variable agents.
• Synergy with quantization/distillation: Chaining with quantization or knowledge distillation for further efficiency.
• Interpretable RLHF/AIF feedback: Mask analysis to identify “which neurons” instantiate which character dimensions, aiding safe and targeted model alignment.
• Digital twin simulation: Applying Persona-Pruner to behavioral simulation in virtual worlds, aligning with directions in digital twin evaluation [li-etal-2025-far].

Conclusion

Persona-Pruner demonstrates that high-character-fidelity role-playing LLMs can be efficiently realized through structured, persona-aligned sub-network extraction alone—without the need for costly per-persona fine-tuning or full model cloning. Its robust numerical performance, strong human concordance, and interpretability position it as a practical and theoretically principled foundation for next-generation, scalable agent ecosystems.