Tag
该论文提出CARE,一种针对视觉-语言-动作(VLA)模型推理加速的认证选择框架,通过配对回放和序贯测试,在95%置信度下保证加速-induced失败风险低于用户设定预算,同时在LIBERO基准上实现9.0–10.8倍加速。
This paper formalizes the problem of when to invoke LLMs in streaming inference systems as a risk-based sequential stopping problem. It proves theoretical guarantees and empirically validates the framework on turbofan degradation data.
This paper introduces a meta-learning framework for anytime-valid certified robustness that uses sequential E-processes to adaptively allocate compute, achieving a 20-fold reduction in sample complexity compared to traditional randomized smoothing while maintaining rigorous statistical guarantees.