DEI: Diversity in Evolutionary Inference for Quality-Diversity Search

A new distributed search framework called DEI uses heterogeneous large language models as mutation operators, achieving substantial gains in quality-diversity search by treating each model’s distinct creative prior as a complementary source of novelty, according to research published on arXiv [1][2]. The framework, formally named Diversity in Evolutionary Inference, assigns different LLMs to peer nodes that communicate through non-blocking collective operations [2]. Unlike homogeneous parallel search, which replicates a single model’s inductive biases across all workers, DEI leverages the differences between models to expand the behavioral repertoire of the search [2]. The approach extends the Digital Red Queen framework: nodes share local optimal solutions at the end of each round to seed the next round’s population, creating cross-model adversarial pressure that the authors argue drives robustness beyond intra-model self-play [2]. Researchers evaluated DEI on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine [2]. A four-node heterogeneous ensemble—comprising GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5—achieved a merged-archive QD-Score of 45.90, compared with 20.46 for a single-node baseline, representing a 124 percent improvement at equal total LLM-call budget [2]. Coverage also rose from 63.0 percent to 80.6 percent of cells, a 28 percent increase [2]. The heterogeneous ensemble further outperformed an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families [2]. The concept of diversity as a driver of search quality echoes principles from evolutionary biology. Sir Ronald Fisher, a founder of population genetics and modern statistics, demonstrated how genetic variation within populations provides the raw material for natural selection [3]. Fisher’s work on quantitative genetics and the design of experiments established mathematical frameworks for understanding how diversity contributes to adaptive outcomes [3]. The DEI results provide what the authors call “the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search” [2]. The findings suggest that combining models with different architectural priors can yield search outcomes that no single model achieves alone [2]. The research appears on arXiv as a preprint and has not yet been peer-reviewed [1].