BILLY: Steering Large Language Models via Merging Persona Vectors for Creative Generation

Basic Information

• Paper ID: 2510.10157
• Title: BILLY: Steering Large Language Models via Merging Persona Vectors for Creative Generation
• Authors: Tsung-Min Pai, Jui-I Wang, Li-Chun Lu, Shao-Hua Sun, Hung-Yi Lee, Kai-Wei Chang
• Classification: cs.CL cs.AI
• Publication Date: October 11, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.10157v1

Abstract

Multi-LLM systems enhance the creativity of large language models by simulating collective human intelligence, but suffer from significant drawbacks including high computational costs and substantial inference latency. To address these limitations, this paper proposes BILLY (BlendIng persona vectors for Large Language model creativitY), a training-free framework that captures the advantages of multi-LLM collaboration within a single model by introducing diverse perspectives and expertise. BILLY operates by extracting and merging multiple distinct persona vectors in the model's activation space, using the merged vector at inference time to guide the model's generation process, thereby achieving multi-perspective outputs without explicit multi-LLM communication.

Research Background and Motivation

Core Problems

1. Computational Burden of Multi-LLM Systems: While traditional multi-LLM collaborative frameworks enhance creativity, they suffer from high computational costs and inference latency issues
2. Limitations of Single-Prompt Methods: Relying solely on prompts to enable models to simultaneously adopt multiple roles and perspectives faces integration difficulties
3. Trade-off Between Efficiency and Creativity: How to significantly reduce computational overhead while maintaining creative advantages

Research Significance

Creativity is widely recognized as a cornerstone of human progress, driving innovation across various domains. Recent research has explored LLM creativity, positioning it as a promising tool for applications such as story writing, design ideation, and scientific discovery. The multi-LLM paradigm simulates collective human intelligence by enabling multiple LLMs to engage in iterative discussions to obtain more comprehensive solutions.

Limitations of Existing Methods

1. Multi-LLM Frameworks: High computational costs, long inference times, and process loss
2. Single-Prompt Methods: Rely on the model's ability to integrate multiple roles, but in practice models may handle each persona separately and struggle to integrate them coherently

Core Contributions

1. Proposes BILLY Framework: A training-free persona vector merging framework that enables multi-perspective creative responses within a single LLM
2. Efficiency and Simplicity: Completely training-free, requiring no additional fine-tuning or multi-LLM communication, significantly reducing computational and token costs
3. Interpretable Mechanism: Operates directly in latent activation space, providing an interpretable creativity control mechanism
4. Performance Improvement: Surpasses single-model prompting and traditional multi-LLM methods on creativity-oriented benchmarks

Methodology Details

Task Definition

Given a creative task input, BILLY aims to generate creative outputs that merge multiple professional perspectives (such as environmentalist, creative professional, futurist) while avoiding the computational overhead of multi-LLM systems.

Model Architecture

1. Persona Vector Extraction (Contrastive Activation)

For persona P, the persona vector is extracted through a contrastive method:

v⃗(l)_P = (1/|D+_P|) ∑(x∈D+_P) a⃗(l)(x) - (1/|D-_P|) ∑(x∈D-_P) a⃗(l)(x)

Where:

• D+_P: Collection of responses exhibiting the persona characteristics
• D-_P: Collection of baseline responses lacking the characteristics
• a⃗(l)(x): Token-averaged residual stream activation at layer l

2. Offline Fusion

Merging N persona vectors into a single composite vector:

v⃗(l)_merged = (1/N) ∑(i=1 to N) v⃗(l)_i

3. Inference-time Steering

Modifying activations during inference:

a⃗(l)_steered = a⃗(l)_original + α · v⃗(l)_merged

Where α is a scaling coefficient, and layer 20 is selected for steering in experiments.

Technical Innovations

1. Activation Space Manipulation: Operating directly at the model's internal representation level rather than relying on external prompt engineering
2. Vector Fusion Strategy: Achieving complex persona combinations through simple averaging of multiple persona vectors
3. Training-Free Design: Adding vectors only at inference time without modifying model weights
4. Reusability: Persona vectors extracted once can be used multiple times, amortizing costs

Experimental Setup

Datasets

Extended benchmarks based on the Wallach-Kogan creativity test, enhanced by Lu et al. using GPT-4:

1. AUT: Requires providing practical innovative uses for everyday objects
2. INSTANCES: Enumerating objects according to specific criteria
3. SIMILARITIES: Identifying connections or similarities between objects
4. SCIENTIFIC: Generating creative solutions for scientific problems

Evaluation Metrics

Adopting two core metrics from the Torrance Tests of Creative Thinking (TTCT):

1. Originality: Assesses the statistical rarity or unconventionality of responses
2. Elaboration: Measures the level of detail and supporting information in responses

Using GPT-4o-mini as the LLM judge with a 1-5 scoring scale.

Comparison Methods

1. Single Agent (SA): Single LLM creative response (temperature 0.7)
2. SA (T=1.0): Single agent with high-temperature decoding (temperature 1.0)
3. SA-MRP: Single agent with multi-role prompting
4. LLM Discussion: Multi-LLM framework with three-stage structured discussion process

Implementation Details

• Base Models: Qwen-2.5-7B-Instruct, Llama-3.1-8B-Instruct, Gemma-3-4B-it
• Steering Layer: Layer 20
• Persona Vectors: Environmentalist, creative professional, futurist, analytical thinker, etc.
• Evaluation: 100 queries per benchmark, computing mean and standard deviation

Experimental Results

Main Results

On four creativity benchmarks, BILLY consistently surpasses all baseline methods on the originality metric:

AUT Task Originality Scores:

• Qwen-2.5-7B: BILLY (4.71) > LLM Discussion (4.24) > SA-MRP (4.07) > SA (4.03)
• Llama-3.1-8B: BILLY (4.38) > LLM Discussion (4.21) > SA-MRP (4.05) > SA (3.73)
• Gemma-3-4B: BILLY (4.99) > SA-MRP (4.96) > SA (4.79) > LLM Discussion (3.70)

Efficiency Comparison:

• Inference Latency: BILLY (19s) vs LLM Discussion (513s), approximately 25x faster
• Token Cost: BILLY reduces token usage by over 95%
• Cost per 10,000 queries: BILLY ($0.30) vs LLM Discussion ($25.50)

Ablation Studies

Persona Vector Combination Analysis:

• A single creative professional vector already demonstrates excellent creative performance
• Increasing from 1 to 3 vectors shows significant improvement
• Further increases from 4-7 vectors yield no significant additional benefits

Case Analysis

Using the "redesign a hospital" task as an example:

• Environmentalist Vector: Generates practical blueprints (e.g., community gardens)
• Creative Professional Vector: Produces artistically rich terminology (e.g., bioluminescent ecological clusters)
• Merged Vector: Achieves true concept fusion, retaining the substantive concepts of the environmentalist while reframing them with the visual style of the creative professional

Activation Projection Analysis

Through analyzing projections of activation changes onto persona vectors:

1. Limitations of Prompting Methods: Multi-role prompting shows positive projection on vCRE but negative on vENV, indicating inability to consistently induce all intended personas
2. BILLY Advantages: Maintains positive projections on both vCRE and vENV, successfully co-activating multiple personas
3. Vector Fusion Effectiveness: Single vectors precisely control corresponding semantic concepts, while merged vectors achieve complex interactions

Related Work

LLM Creativity Enhancement

1. Role-Playing Prompting: Guiding models to adopt specific creative personas
2. Multi-Agent Collaborative Frameworks: Multiple agents simulating brainstorming sessions
3. Comprehensive Reviews: Thorough surveys of creativity research directions

Multi-LLM Collaboration

1. Structured Dialogue Frameworks: Such as next-generation applications of AutoGen
2. Role-Playing Dialogue: LLM Discussion and HOLLMWOOD
3. Limitations: High computational costs, instability, and inefficiency

Activation Steering Techniques

1. Behavioral Control: Influencing model behavior through modifying internal activations
2. Contrastive Methods: Computing steering vectors using activation differences
3. Application Domains: Authenticity enhancement, safety improvement, and personality control

Conclusions and Discussion

Main Conclusions

1. BILLY successfully captures the creative advantages of multi-LLM collaboration within a single model
2. Significantly reduces inference time and computational costs while maintaining or exceeding creative performance
3. Activation steering provides better controllability and interpretability compared to multi-role prompting

Limitations

1. Vector Combination Method: Currently uses simple averaging; more complex combination techniques are needed in the future
2. Non-linear Interactions: The fusion effects of persona vectors are not simple linear superposition
3. Task-Specific Weighting: Lacks mechanisms for learning task-specific weights for each persona vector

Future Directions

1. Develop complex vector combination frameworks beyond simple averaging
2. Learn task-specific weights for persona vectors
3. Design mechanisms that explicitly model functional roles for more precise control

In-Depth Evaluation

Strengths

1. Methodological Innovation: First application of persona vector merging to creativity enhancement with a novel technical approach
2. Comprehensive Experiments: Thorough evaluation across multiple models and benchmarks, including efficiency, quality, and interpretability analysis
3. Practical Value: Significant efficiency improvements make the method highly practical
4. Theoretical Insights: Activation projection analysis reveals the fundamental differences between prompting methods and activation steering

Weaknesses

1. Vector Fusion Theory: Lacks in-depth theoretical analysis of why simple averaging is effective
2. Evaluation Limitations: Primarily relies on LLM judges, which may introduce bias
3. Generalization Capability: Validated only on creativity tasks; applicability to other tasks remains unknown
4. Persona Vector Quality: Stability and consistency of the persona vector extraction process require further verification

Impact

1. Domain Contribution: Provides a new technical pathway for LLM creativity enhancement
2. Practical Value: Substantially reduces computational costs, promising to facilitate practical deployment of multi-persona AI systems
3. Reproducibility: The method is relatively simple and easy to reproduce and extend

Applicable Scenarios

1. Creative Writing: Story creation requiring fusion of multiple perspectives
2. Design Ideation: Design solutions combining aesthetics, functionality, and sustainability
3. Scientific Innovation: Research problem-solving requiring interdisciplinary thinking
4. Educational Applications: Cultivating students' multi-perspective thinking abilities

References

The paper cites abundant related work, including:

• Classical creativity psychology research (Torrance, 1966; Wallach & Kogan, 1965)
• Recent advances in multi-agent systems (Wu et al., 2023; Lu et al., 2024)
• Development of activation steering techniques (Turner et al., 2024; Chen et al., 2025)

Overall Assessment: This is a high-quality paper with significant contributions to the field of LLM creativity enhancement. The BILLY method ingeniously addresses the efficiency problems of multi-LLM systems while maintaining creative advantages. Although there is room for improvement in theoretical analysis and evaluation methods, its practical value and technical innovation make it an important advance in the field.