Technical Summary: BILLY (BlendIng persona vectors for Large Language model creativitY)

Problem Statement

While multi-LLM systems have demonstrated potential in enhancing creativity by simulating human collective intelligence through iterative discussions, they suffer from significant practical limitations. These include high computational costs, increased inference latency due to multi-round token exchanges, and "process loss," where coordination inefficiencies prevent the collective output from surpassing the sum of individual contributions. Conversely, single-model approaches relying on prompt engineering (e.g., asking a model to adopt multiple roles simultaneously) often struggle to coherently integrate diverse perspectives within a single response, leading to inconsistent or superficial results. The core challenge addressed is how to simulate diverse, multi-perspective collaboration within a single LLM efficiently, without the overhead of multi-agent communication or the need for additional model training.

Methodology

The authors propose BILLY, a training-free framework that steers a single Large Language Model (LLM) by blending distinct persona vectors directly within the model's activation space. The approach consists of three primary stages:

1. Extraction of Persona Vectors:
   Using a contrastive activation method, the authors extract directional vectors representing specific personas (e.g., "Creative Professional," "Environmentalist"). This involves curating two datasets for each persona $P$: a positive-expression set ($D^+_P$) where the model exhibits the persona's traits, and a negative-expression set ($D^-_P$) containing neutral or baseline responses. The persona vector $\vec{v}^{(l)}_P$ at layer $l$ is calculated as the difference between the mean token-averaged residual stream activations of these two sets:
   $$\vec{v}^{(l)}_P = \frac{1}{|D^+_P|} \sum_{x \in D^+_P} \vec{a}^{(l)}(x) - \frac{1}{|D^-_P|} \sum_{x \in D^-_P} \vec{a}^{(l)}(x)$$

2. Offline Fusion:
   Unlike multi-agent systems that require online interaction, BILLY fuses multiple extracted persona vectors offline into a single composite steering vector ($\vec{v}_{merged}$). This is achieved by simply averaging the individual vectors:
   $$\vec{v}^{(l)}_{merged} = \frac{1}{N} \sum_{i=1}^{N} \vec{v}^{(l)}_i$$
   This composite vector encapsulates a multi-faceted perspective in the latent space.

3. Inference-Time Steering:
   During generation, the composite vector is added to the model's activations at a specific layer $l$ (experimentally set to layer 20) with a scaling coefficient $\alpha$ (experimentally set to 2.0):
   $$\vec{a}^{(l)}_{steered} = \vec{a}^{(l)}_{original} + \alpha \cdot \vec{v}^{(l)}_{merged}$$
   This operation nudges the model's internal state toward the fused perspectives without altering the model's weights or requiring additional training.

Key Contributions

The paper outlines three primary contributions:

• Enhanced Creativity: BILLY enables the generation of diverse, multi-perspective responses that surpass both single-model prompting and traditional multi-LLM approaches on creativity-oriented benchmarks.
• Efficiency and Simplicity: The framework is entirely training-free and eliminates the need for multi-LLM communication. It achieves comparable or superior creative performance with substantially lower computational and token costs.
• Interpretability: By operating in the latent activation space, BILLY provides a transparent mechanism for steering creativity. The authors demonstrate that distinct persona vectors can be blended to control complementary aspects of generation (e.g., style vs. content) with fine-grained control.

Experimental Results

The authors evaluated BILLY on four creativity benchmarks derived from the Wallach-Kogan Creativity Tests: AUT (Alternative Uses), INSTANCES, SIMILARITIES, and SCIENTIFIC. The evaluation utilized three base models (Qwen-2.5-7B, Llama-3.1-8B, and Gemma-3-4B) and measured performance using Originality and Elaboration scores, assessed via both LLM-based judges (GPT-4o-mini) and human evaluators.

• Performance: BILLY consistently achieved higher Originality scores than all baselines, including Single Agent (SA), SA with high-temperature decoding, SA with Multi-Role Prompting (SA-MRP), and the multi-agent "LLM Discussion" framework. It maintained competitive or superior Elaboration scores.
• Efficiency: BILLY demonstrated a dramatic reduction in inference time (approximately 25x faster than LLM Discussion) and reduced token costs by over 95% compared to multi-agent baselines.
• Analysis:
  • Vector Composition: Merging vectors showed diminishing returns beyond a certain number of personas (e.g., 3 to 4 vectors), suggesting the benefit is not purely additive.
  • Steering Alignment: Projection analysis revealed that while prompt-based multi-role prompting (SA-MRP) failed to consistently activate all intended personas in the latent space, BILLY successfully co-activated multiple personas, validating its superior controllability.
  • Qualitative Fusion: Case studies indicated that the merged vector successfully synthesized the substantive content of one persona (e.g., Environmentalist) with the stylistic flair of another (e.g., Creative Professional), whereas prompt-based methods often resulted in analytical or disjointed outputs.

Significance and Claims

The paper claims that BILLY offers a practical solution for real-time creative generation by capturing the benefits of multi-LLM collaboration within a single model. The authors argue that their method demonstrates that the creative advantages of multi-agent systems can be effectively captured via activation steering, significantly reducing the inference time and token costs associated with creativity tasks. Furthermore, the work suggests that activation steering with composite persona vectors offers superior controllability and interpretability compared to multiple role-play prompting, as it directly manipulates the model's internal state to align with complex, multi-faceted personas. The authors acknowledge that while their current method uses simple averaging for vector fusion, future work could explore more sophisticated composition techniques to better manage non-linear interactions between personas.