Does Competition-Aware prompting elicits higher performance from Large Language Models?

Large language models are sensitive to small inference-time prompt interventions, but most methods designed to improve performance focus on reasoning instructions, demonstrations, sampling, or additional test-time computation. We study whether contextual information that is not itself a solution strategy can affect performance by asking a different question: does performance change when the prompt states that the model is competing against other AI models? We introduce competition-aware prompting and evaluate it across three open-weight language models and four knowledge and reasoning benchmarks with placebo and paraphrase control. The pooled C2–C1 effect was +0.02 pp, with practical equivalence within ±2 pp. A provisional Vicuna 13B extension showed a promising MMLU-Pro gain of +4.52 pp (95% CI: +1.43 to +7.86), although its overall confidence interval included zero. This benchmark-specific signal warrants further investigation of when competition-aware prompting may improve performance.