Anchorless Diversification for Parallel LLM Ideation | AI Research

Anchorless Diversification for Parallel LLM Ideation
Large Language Models (LLMs) are frequently used to generate pools of creative ideas, such as story premises or product concepts. A common challenge is that these models often produce similar, redundant outputs because they tend to gravitate toward familiar patterns. While researchers have developed ways to force diversity—such as asking the model to critique its own work or comparing ideas against previous outputs—these methods often require extra computational steps, shared state, or multiple rounds of generation, which can be inefficient. This paper explores "anchorless" methods for diversification, testing whether we can achieve a broad, high-quality pool of ideas without needing to observe seed examples or coordinate between multiple generation steps.

The Power of Direct Instructions

The researchers tested a simple, low-cost strategy: using a "population-referential" instruction. Instead of just asking the model to be creative, they instructed it to generate a response that stands out from other potential responses for the same task. This approach does not require the model to see any prior examples; it relies on the model’s internal knowledge of what a typical response looks like. The results show that this simple instruction consistently improves the variety of the idea pool while maintaining or even enhancing the quality of the outputs, all while keeping computational costs very low.

Organizing Ideas with Semantic Stratification

A more advanced anchorless technique introduced in the paper is "semantic direction stratification." In this method, the system performs a single initial planning call to identify broad, distinct semantic categories for the task. The model then allocates its generation budget evenly across these categories. By mapping out the "creative space" before generating the final pool, the model can spread its ideas across different regions effectively. This method proved to be highly efficient, offering one of the best balances between diversity, quality, and computational cost.

Comparing Efficiency and Performance

The study compared these anchorless methods against "anchored" approaches, which rely on observing seed ideas or peer outputs to guide subsequent generations. While anchored methods can be effective at creating diverse pools, their advantage often diminishes when the full cost of the entire generation pipeline—including the time and tokens spent creating and processing seed ideas—is taken into account. The researchers found that combining population-referential instructions with semantic stratification provides the strongest performance, establishing these as practical, cost-effective baselines for anyone looking to generate diverse creative pools using LLMs.