Concept-Guided Prompting Strategies Increase Diversity in LLM-Generated Recipe Populations

Stuart FarmerarXiv 2026 · arXiv

TL;DR

LLMs often struggle with "repetitive output syndrome" when asked to generate large sets of solutions. This paper introduces concept-guided prompting, a hierarchical strategy that first generates diverse thematic concepts and then populates them with specific solutions, significantly increasing the semantic diversity and human-like coverage of generated populations without the need for expensive iterative feedback loops.

What Problem Does This Paper Address?

When using Large Language Models (LLMs) for automated discovery or population-based optimization (like generating a library of code snippets or a book of recipes), the models tend to converge on a narrow set of "safe" or high-probability responses. This lack of diversity—getting stuck in local minima—is a major hurdle for systems that rely on exploring a wide "behavior space."

While existing techniques like Quality-Diversity (QD) algorithms (e.g., MAP-Elites) address this by explicitly measuring and rewarding novelty, they often require an "evaluator" LLM to score every generated item, which is computationally expensive and slow. The author asks: Can we encourage broader populations through prompting structure alone, during the generation phase itself?

Methods

The author proposes concept-guided prompting, which decomposes a generation task into a two-layer tree structure:

  1. The Concept Layer: The LLM is first asked to generate a batch of diverse thematic "concepts" (e.g., "Spicy Fusion," "Minimalist Baking").
  2. The Solution Layer: For each concept, the LLM generates a small batch of specific solutions (recipes) steered by that theme.
Figure 2: A tree of height two connects one problem context to three concepts, each guiding three solu- tions.

This approach was compared against two baseline strategies:

  • Nonces: Repeatedly querying the LLM with a small random string (nonce) added to the prompt to jitter the model's internal state.
  • Prior Titles: Generating recipes sequentially and feeding the list of previously generated titles back into the prompt with instructions to "avoid these."

To evaluate diversity, the author used text-embedding-3-large to map all recipes into a 3,072-dimensional vector space, using geometric measures as a proxy for semantic variety.

Datasets

  • Generated Populations: 1,000 recipes were generated for each of the three strategies (3,000 total) using the gpt-oss-120b:nitro model.
  • Human Reference Corpus: A control group of 168 human-authored recipes from "Text Only Recipes" was used to define the boundaries of "useful diversity"—ensuring the generated recipes weren't just diverse, but occupied a semantically valid "culinary space."

Key Insights and Ideas

  • Semantic Islands: By forcing the model to commit to a theme first, the strategy creates "semantic islands." This prevents the model from drifting back toward its most probable (and repetitive) outputs, effectively "steering" it into under-explored regions of the behavior space.
  • Participation Ratio as a Diversity Metric: The paper uses the "participation ratio" to measure how many principal directions of the embedding space a population occupies. A higher ratio means the variation is spread across many semantic dimensions rather than just a few.
  • The "Human-to-Generated" Flip: A fascinating finding was that while individual concept-guided recipes were often further from their nearest human neighbor than the baselines, the entire population covered more of the human "neighborhoods." This suggests the method produces a "shotgun blast" that covers a wider area, even if individual pellets don't hit the bullseye of existing human examples.

Key Results

Concept-guided prompting outperformed both baselines across nearly every metric:

  • Semantic Coverage: It reached 47.0% coverage of human recipe neighborhoods, compared to 35.1% for the title-avoidance strategy and only 13.1% for the nonce strategy.
  • Internal Spacing: The "local spacing" (average distance to the 15 nearest neighbors) for concept-guided recipes was 1.59x higher than the title-avoidance baseline and 15x higher than the nonce baseline.
  • Distribution of Diversity: As shown in the density plots, the concept-guided strategy shifted the entire population toward higher mean squared neighbor distances, indicating a consistently more diverse set.
Figure 6b: Gaussian kernel density estimates of local spac- ing scores, with Scott’s bandwidth per population and reflection at zero. Each curve integrates to one; vertical dashed lines show observed means.

Limitations and Future Work

  • Quality vs. Diversity: The study focused entirely on semantic diversity using embeddings as a proxy. It did not evaluate the "culinary validity" or quality of the recipes. A highly diverse set of recipes is useless if they are all inedible.
  • Generalizability: The experiment was limited to a single "toy task" (recipes). Further research is needed to see if this hierarchical steering works for more complex domains like code synthesis or mathematical theorem proving.
  • Concept Bottleneck: The diversity of the final population is heavily dependent on the diversity of the initial "concept layer." If the LLM generates 100 very similar concepts, the resulting 1,000 recipes will likely lack breadth.

Summary updated

Abstract

Novelty in solution generation supports automated discovery, but LLM outputs can remain repetitive. To address this, we propose a simple concept-guided prompting method that generates thematic concepts before responses within each of these concepts. To demonstrate, we generated 1,000 recipes for a recipe book task and compared the results against a human-authored corpus. Two naive prompting strategies were compared as baselines. Concept-guided prompting outperformed both baselines in all cases and reached semantic coverage of 47.0% versus 35.1% and 13.1% in respective prior-title and nonce prompt strategies. These findings suggest broader population exploration and diversity, but remain contained to this specific task. Further work is needed to determine the quality of solutions and generalizability.
BibTeX
@misc{farmer2026concept,
  author = {Stuart Farmer},
  title = {{Concept-Guided Prompting Strategies Increase Diversity in LLM-Generated Recipe Populations}},
  year = {2026},
  eprint = {2611.01171},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.01171},
}
Figure