Higher 'generative-process diversity' between a model pair predicts lower correlated failure, beyond semantic similarity or capability — a single, not-yet-corroborated 2026 preprint

ai-safetycorrelated-errorsalgorithmic-monoculturerepresentational-similaritynlp-evaluation

claim-correlated-errors-in-llms-exceed-chance-and-rise-with-accuracy establishes that LLM errors correlate above chance; this preprint proposes and tests a candidate mechanism. It defines "generative-process diversity" — how different the underlying processes capable of producing two models' outputs are, estimated via compression-distance between raw outputs rather than by comparing the outputs' meaning — and finds, across 38 models and ten disjoint benchmark families, that higher generative-process diversity between a model pair predicts lower correlated failure, beyond what semantic similarity or individual capability explain (partial rank association −0.216, negative on all ten benchmarks tested). The paper frames the broader pattern it is building on plainly: "Models from different frontier labs select the same wrong answers more often than expected from their individual error distributions." [unverified-quant — single, very recent preprint, not yet independently corroborated; see question-corroborate-tieman-markou-2026-generative-process-diversity].