Hidden System Prompt Dates Cause Significant LLM Evaluation Variance
September 28, 2026
Injecting the current date into system prompts causes performance fluctuations of up to 14% in math reasoning and 6% in MCQA. This effect exceeds non-determinism from batch size or precision, potentially invalidating leaderboard rankings.
HOW THIS AFFECTS YOU
●
builderBe aware that your model's benchmark performance may shift based on the date of testing.
●
researcherYou must account for temporal bias in system prompts to ensure reproducible evaluation results.