Abstract
Infrastructure changes in LLM serving can alter functional outputs even when model weights and prompts remain fixed. DriftBench measures these changes across hardware, precision, framework, model, and workload dimensions, then evaluates when risk can be predicted from configuration metadata. Its Portability Risk Index transfers strongly to unseen hardware and precision but not to every framework or model change, yielding a practical rule: predict systematic transitions and re-measure idiosyncratic ones. The study also shows why aggregate quality metrics are insufficient by identifying prompt-level safety and correctness flips during infrastructure migration. The contribution is a measurement and risk-assessment framework, not a drift-mitigation method.
Citation
G. Vitale, “DriftBench: Measuring and Predicting Infrastructure Drift in LLM Serving Systems,” Proceedings of Machine Learning and Systems 8, 2026.