Researchers have tested whether large language models can serve as a cheap substitute for the expensive engineering practice of building redundant, independently written software.
The classic approach to reducing common-mode failures in critical systems is software diversity: pay multiple teams to implement the same specification, then run them in parallel and compare outputs. It works, but it's expensive to build, validate, and maintain. The new paper extends that classical research to LLM-generated code, testing three specifications against large pools of programs produced by different models, temperatures, and programming languages. Using a 1-out-of-2 voting configuration — where a system passes if at least one of two implementations gives the correct answer — the researchers found that pairing LLM-generated programs in heterogeneous combinations does yield measurable reliability gains over any single implementation.
The practical implication is meaningful: if LLMs can cheaply generate a diverse enough population of implementations, the cost barrier that kept software redundancy out of most projects drops significantly. The catch is that gains are uneven — they depend on which programming language is used and how the model's generation temperature is set, meaning there is no simple plug-and-play answer.
For all the enthusiasm around AI-generated code, most of the conversation has focused on developer productivity. This research points at a different value: not just writing code faster, but using statistical variation across many generated versions to harden what ships. Whether that holds up in production systems more complex than the three specifications tested here is the question this paper leaves open.