LLMEffiChecker: Understanding and Testing Efficiency Degradation of Large Language Models

Xiaoning Feng, Xiaohong Han, Simin Chen, Wei Yang

Summary

LLMEffiChecker examines the efficiency robustness of large language models. It generates subtle input perturbations that delay end-of-sentence generation and increase the amount of inference computation. The approach supports gradient-guided white-box testing and causal-inference-based black-box testing. Evaluation on nine models demonstrates substantial increases in latency and energy consumption, including effects on mobile-device battery usage.