Preprint finds AI code optimization usually increases execution energy

A preprint testing 11 open-source models on all 1,000 EffiBench problems found that 42 of 44 configurations produced code consuming more energy per problem than canonical human reference solutions. The excess ranged from +4% to +886%. Researchers measured energy during iterative model refinement and code execution. Limiting maximum output tokens reduced inference energy by up to 69% against the default configuration. Mixture-of-Experts models used substantially less inference energy than equivalently-sized dense models without sacrificing correctness. The two configurations that improved execution energy achieved only marginal gains, requiring on the order of 10^6 to 10^7 executions to offset optimization energy.