CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse
CreativityNeuro enhances divergent thinking by contrastive weight steering, improving DAT performance by 14 percentile points.
Key Findings
Methodology
CreativityNeuro uses contrastive weight steering without relying on datasets or gradient-based fine-tuning. By contrasting creative and non-creative prompts, it computes parameter importance, selects creativity-relevant parameters, and applies weight perturbations.
Key Results
- CreativityNeuro improved DAT performance by 14 human percentile points, significantly outperforming other baselines.
- In AUT and Task Task, CreativityNeuro enhanced originality and surprise, outperforming activation steering.
- CreativityNeuro reduced mode collapse across all tasks.
Significance
This study provides a simple method to enhance creativity in large language models without behavioral data or retraining, addressing the issue of generating similar responses in open-ended tasks.
Technical Contribution
CreativityNeuro performs adjustments in weight space, avoiding dependency on behavioral data, and shows good generalization to unseen tasks.
Novelty
First to enhance divergent thinking without data, through weight steering, differing from activation steering methods reliant on behavioral data.
Limitations
- In some complex tasks, creativity enhancement may lead to decreased utility.
- The method's effectiveness may vary across different models.
Future Work
Future research could explore enhancing creativity without affecting utility and validate the method's applicability across more tasks.
AI Executive Summary
Large language models often generate similar responses to open-ended questions, limiting creativity. Existing methods rely on behavioral data or gradient fine-tuning, which are hard to apply broadly. CreativityNeuro enhances divergent thinking through contrastive weight steering without datasets or retraining.
CreativityNeuro excels in the Divergent Association Task (DAT), improving by 14 human percentile points. In the Alternative Uses Test (AUT) and Task Task, it significantly boosts originality and surprise, while reducing mode collapse. Compared to activation steering, CreativityNeuro shows stronger generalization capabilities.
While creativity enhancement may affect utility in some tasks, CreativityNeuro offers a new approach to improving model performance in creative domains. Future research could further optimize the method to enhance creativity without compromising utility.
Deep Analysis
Background
Recent advances in large language models have shown significant progress in generative tasks, but they often suffer from mode collapse when handling open-ended questions, limiting creativity. Existing methods rely heavily on behavioral data or gradient fine-tuning, limiting their broad applicability.
Core Problem
The issue of large language models generating similar responses to open-ended questions is known as the artificial hivemind effect. This limits their application in creative tasks, especially where diverse and novel solutions are required.
Innovation
CreativityNeuro enhances divergent thinking through contrastive weight steering. Unlike activation steering, which relies on behavioral data, this method requires no datasets or retraining, demonstrating good generalization to unseen tasks.
Methodology
- �� Compute parameter importance by contrasting creative and non-creative prompts
- �� Select parameters relevant to creativity
- �� Apply weight perturbations to enhance model creativity
Experiments
Experiments were conducted on DAT, AUT, and Task Task using various models and prompt sets. Human evaluations confirmed improvements in originality and surprise with CreativityNeuro.
Results
CreativityNeuro improved DAT performance by 14 human percentile points, enhanced originality and surprise in AUT and Task Task, and reduced mode collapse.
Applications
The method can be used to enhance large language models' performance in creative tasks, such as creative writing and advertising copy generation.
Limitations & Outlook
While the method enhances creativity, it may decrease utility in some tasks. Future research could explore how to enhance creativity without affecting utility.
Plain Language Accessible to non-experts
Imagine you're in a kitchen cooking. Existing large language models are like a chef who can only cook one dish, skilled but lacking creativity. CreativityNeuro is like giving the chef a new cookbook, allowing them to try different recipes. By adjusting some key ingredient ratios (i.e., model weights), the chef can create more diverse dishes without relearning basic cooking skills. This way, the chef can not only make delicious dishes but also innovate continuously to satisfy different diners' tastes.
ELI14 Explained like you're 14
Imagine you're playing a game where the characters always do the same thing. Wouldn't that be boring? CreativityNeuro is like giving the characters new skills, allowing them to do more interesting actions! This makes the game more fun! And these new skills are achieved by adjusting the characters' attributes, without redesigning them! It's like giving your game character new gear, offering more choices and possibilities in their adventures!
Glossary
Divergent Thinking
The ability to generate diverse and novel solutions, often used to assess creativity.
Used in the paper to evaluate model performance in open-ended tasks.
Mode Collapse
The tendency for models to generate similar responses, lacking diversity and novelty.
Limits creativity in open-ended tasks.
Contrastive Weight Steering
A method that adjusts model weights by contrasting different prompts to enhance specific abilities.
CreativityNeuro uses this method to enhance divergent thinking.
Divergent Association Task (DAT)
A task assessing divergent thinking by generating semantically distant words.
CreativityNeuro excels in this task.
Alternative Uses Test (AUT)
A psychometric tool requiring creative uses for common objects.
Used to evaluate CreativityNeuro's performance in complex tasks.
Open Questions Unanswered questions from this research
- 1 How to enhance creativity without affecting utility? Current methods may decrease utility in some tasks, requiring new optimization strategies.
- 2 Does CreativityNeuro perform consistently across different models and tasks? More experiments are needed to validate its generalization capabilities.
Applications
Immediate Applications
Creative Writing
Enhance large language models' performance in creative writing, generating more novel and diverse content.
Long-term Vision
Intelligent Creative Assistant
Achieve more intelligent creative assistants to provide inspiration and suggestions in various creative tasks.
Abstract
Divergent thinking is a crucial aspect of creativity, yet large language models (LLMs) tend to consistently generate similar responses to open-ended questions, in what has been termed the artificial hivemind effect. Here, we introduce CreativityNeuro, a data-free method for enhancing divergent thinking in LLMs via contrastive weight steering. We evaluate our method across multiple creativity assessments and report several main findings. On the Divergent Association Task (DAT), a vocabulary-space creativity test, CreativityNeuro improves performance by up to 14 human percentile points. Next, in a large-scale human evaluation (N=720) on the Alternative Uses Test (AUT) and the Task Task, CreativityNeuro achieves significant improvements in originality, surprise, and creativity, transferring to longer-form and more open-ended tasks. Importantly, we find that across all three tasks, CreativityNeuro demonstrably reduces measures of mode collapse. Moreover, activation steering achieves comparable performance to CreativityNeuro on the DAT, but it does not transfer to the AUT and Task Task, demonstrating the effectiveness of weight-space steering in generalizing to unseen tasks. In conclusion, CreativityNeuro improves divergent thinking and reduces mode collapse without requiring behavioral data, re-training, or gradient-based fine-tuning, providing a straightforward way to enhance LLM performance in creative domains.