Breaking LLM Statistical Convergence: A Bisociative Prompt Engineering Toolkit

The open-source toolkit bisociative-ai-creative-prompting helps developers break through the bottleneck of homogeneous output from large language models (LLMs) using bisociative prompt strategies and real-time similarity analysis. This article will discuss the predicament of homogeneous LLM output, the theoretical basis of bisociative thinking, the core components of the toolkit, technical implementation, application value, and future directions.

With the widespread application of LLMs, the problem of homogeneous output has become increasingly prominent—mainstream models produce highly consistent response patterns and wording for similar questions. The root cause lies in the training mechanism of LLMs, which predicts the next word based on probability and tends to choose the most common expressions from training data rather than creative or unique insights. In scenarios like creative writing and brainstorming, this "statistical convergence" becomes a shortcoming, making it difficult to meet users' needs for novel perspectives.

The toolkit draws inspiration from psychologist Arthur Koestler's concept of "bisociation"—creative thinking arises from the collision of two unrelated thinking frameworks. Based on this, the toolkit designs prompt strategies to artificially create framework collisions and break the inherent statistical convergence tendency of LLMs.

Convergence Analyzer

Quantitatively evaluates homogeneous LLM output. Its functions include: measuring the semantic similarity of multiple generated results, identifying the model's "comfort zone" expression patterns, and quantifying the impact of prompt strategies on diversity, serving as an optimization feedback loop.

Five Bisociative Prompt Strategies

1. Domain Grafting: Map problems across domains (e.g., software development → culinary terms);
2. Role Conflict: Activate multi-role debates (e.g., auditor vs. entrepreneur);
3. Temporal Dislocation: Place in historical/future contexts (perspective from the Renaissance or 500 years later);
4. Scale Transition: Switch between micro and macro scales (urban traffic → planetary network / single intersection);
5. Negation Reconstruction: First list impossible methods then find inspiration.

The toolkit is a Python library with a modular design for easy integration. The convergence analyzer is based on sentence-transformers and cosine similarity. Usage flow: 1. Generate baseline output with default prompts and measure convergence; 2. Regenerate using bisociative strategies; 3. Compare diversity improvements. Jupyter Notebook examples are provided to demonstrate the full process.

Directly useful for creative scenarios: content creators break through writing bottlenecks, product managers conduct comprehensive brainstorming, and researchers analyze from multiple angles. More importantly, it provides a systematic method to combat LLM statistical convergence, offering a feasible direction for convergence issues after model scale expansion.

Bisociative prompting is not a panacea: over-pursuing diversity may lose coherence/practicality, and convergence is an advantage in precise scenarios (code generation, factual Q&A), so strategies need to be chosen flexibly. Future plans include introducing an automatic strategy selection mechanism to lower the usage threshold.