Technique that applies multiple transformations at inference and aggregates outputs to improve robustness, common in vision and sometimes LLM prompts.