Diversity Pressure at the Capability Ceiling: A 50-Run Controlled Study Showing No Measurable Benefit on Ceiling-Difficulty Targets

Read full text Summarise with VIRAT
Authors: Blake Musselman
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Diversity-forcing mechanisms in LLM code generation — systems that apply iterative mutation and selection pressure to produce structurally distinct candidate solutions — have demonstrated measurable benefits on problems at the model's capability boundary. This paper reports a negative result from a 50-run controlled experiment comparing evolutionary diversity pressure against single-shot baseline generation on ceiling-difficulty targets: problems that exceed the model's reliable capability. The verdict is unambiguous: across all 50 paired comparisons, evolutionary generation and single-shot baseline produce statistically indistinguishable outcomes. Both strategies fail at comparable rates; neither produces viable solutions the other cannot. This is a ceiling effect: diversity pressure selects among the solutions a model can produce, not from solutions the model cannot produce. The finding refines an earlier positive result by partitioning problem complexity into two regimes — boundary complexity, where diversity forcing is demonstrably effective, and ceiling complexity, where it is not. The practical implication: diversity pressure is not a universal enhancement. It is a tool calibrated for the boundary zone.

Fields of Study & Subjects

Evolutionary Algorithms and Applications Software Engineering Research Metaheuristic Optimization Algorithms Research

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
EN
Document type
Preprint