Technical Summary: CreativityNeuro

Problem Statement

Large Language Models (LLMs) exhibit a phenomenon termed the "artificial hivemind effect," where they consistently generate similar, homogenized responses to open-ended questions. This behavior limits their capacity for divergent thinking—the ability to generate multiple diverse solutions to a problem—which is a crucial component of creativity. Existing methods to enhance creativity, such as prompting, decoding parameter tuning (e.g., temperature), and activation steering, often require behavioral datasets, gradient-based fine-tuning, or fail to generalize across different creative tasks. Furthermore, there is a lack of understanding regarding whether creative behaviors are separable from factual reasoning within the model's weight space.

Methodology: CreativityNeuro

The authors introduce CreativityNeuro (CN), a data-free method for steering LLM weights to amplify divergent thinking. Unlike previous approaches that rely on behavioral data or gradient updates, CN operates by identifying and scaling specific parameters in the model's weight space based on contrastive prompt sets.

Core Algorithm

The method follows a three-step process (detailed in Algorithm 1):

1. Contrastive Prompt Sets: Instead of relying on structured datasets (like MATH for reasoning), CN utilizes six sets of contrastive prompts ($P_{cre}$ vs. $P_{non-cre}$) spanning styles such as storytelling, ideation, and minimal instructions (e.g., "Surprise me" vs. "Be precise").
2. Parameter Importance Scoring: Using a Wanda-style scoring mechanism, the method computes importance scores ($S_{\ell,ij}$) for weights in each layer $\ell$. The score is the product of the weight magnitude and the activation norm, summed across tokens in the prompt sets:
   $$S_{\ell,ij} = \sum_{b,t} |W_{\ell,ij}| \cdot \|x^{(b,t)}_{\ell,j}\|^2$$
   Scores are calculated separately for creative ($P_{cre}$) and non-creative ($P_{non-cre}$) prompts.
3. Weight Selection and Scaling: The method isolates creativity-specific parameters by selecting the top $\rho$ percent of weights ranked by creative importance that do not appear in the top $\rho$ percent for non-creative prompts ($C_\ell \setminus N_\ell$). These selected weights are then scaled by a factor $(1 + \alpha)$ during inference.

This approach requires no behavioral generations, scored responses, or labeled examples, making it entirely data-free.

Key Contributions

The paper outlines five primary contributions:

1. Data-Free Steering: Introduction of CreativityNeuro, a method to steer creative behavior without behavioral datasets or retraining.
2. Divergent Association Task (DAT) Improvement: Demonstration that CN significantly improves DAT scores, outperforming baselines including prompting, sampling parameter variations, and activation steering.
3. Generalization to Complex Tasks: Validation via large-scale human evaluation ($N=720$) on the Alternative Uses Test (AUT) and Task Task (TT), showing improvements in originality, surprise, and creativity.
4. Mode Collapse Reduction: Evidence that CN reduces both word-level (DAT) and response-level (AUT/TT) mode collapse.
5. Weight Space Entanglement: Discovery that divergent thinking and factual reasoning are non-separable in weight space; attempts to isolate them degrade performance in both domains.

Experimental Results

1. Divergent Association Task (DAT)

Evaluated across six instruct-tuned models (Phi, Llama, Qwen families, ranging from 3B to 14B parameters):

• Performance: CN improved DAT performance by up to 14 human percentile points.
• Comparison: CN achieved an average human percentile of 94.1, slightly outperforming activation steering (93.9) and significantly outperforming prompting and sampling baselines.
• Data Efficiency: Unlike activation steering, which requires scored behavioral data to construct steering vectors, CN achieved superior results using only prompt pairs.

2. Alternative Uses Test (AUT) and Task Task (TT)

A large-scale human evaluation ($N=720$) assessed originality, surprise, utility, and creativity:

• AUT: CN achieved uniformly positive effects on originality (average Cohen's $d = +0.36$) and surprise ($d = +0.43$).
• TT: CN showed strong gains in originality ($d = +0.40$) and moderate improvements in creativity ($d = +0.24$).
• Generalization: While activation steering performed comparably to CN on the DAT, it failed to generalize to the AUT and TT. This highlights the superiority of weight-space steering over activation steering for unseen, complex creative tasks.
• Trade-offs: CN reduced utility scores on the AUT, a predictable consequence of the novelty-utility tradeoff inherent in creative generation.

3. Mode Collapse

• DAT: CN reduced the probability mass on the top 10 most frequent words by 10.2 percentage points and increased vocabulary entropy by 10%.
• AUT/TT: CN reduced intra-model repetition and inter-model homogeneity, demonstrating a reduction in semantic mode collapse across sentence and paragraph-length responses.

4. Separability of Divergent Thinking and Factual Reasoning

The authors investigated whether creative weights could be isolated from factual reasoning weights (tested on the MMLU benchmark):

• Finding: Divergent thinking and factual reasoning are functionally entangled in weight space.
• Evidence: Attempts to protect MMLU performance by excluding MMLU-related weights from the creative mask resulted in further degradation of MMLU scores, despite reducing the mask size. This suggests that the parameters driving creativity are deeply intertwined with those driving factual accuracy.

Significance and Claims

The paper claims that CreativityNeuro provides a straightforward, data-free mechanism to enhance LLM performance in creative domains. Its primary significance lies in:

• Generalization: Proving that weight-space steering generalizes to unseen creative tasks (AUT, TT) where activation steering fails.
• Efficiency: Achieving state-of-the-art creative performance without the computational cost or data requirements of fine-tuning or behavioral dataset collection.
• Mechanistic Insight: Offering evidence that creativity and factual reasoning are not modularly separable in current LLM architectures, suggesting that enhancing one may inherently impact the other.

The authors conclude that while CN effectively improves divergent thinking and reduces homogenization, the entanglement of creative and factual weights remains a fundamental architectural constraint that future research must address, potentially through multi-agent systems or unified factored representations.