Conformal Cache: Reliable Proxy-Discrepancy Caching for Fast Generative Inference
Published in Neural Information Processing Systems (NeurIPS) , 2026 [CCF A]
Wenxiao Wu, Yachen Gao, Yukang Feng, Chengming Xu, Moran Li, Jingyu Li, Ming Xie, Xiaobin Hu, Xinwei Sun, Jing-Hao Xue, Nong Sang, Yanwei Fu
Conformal Cache develops a reliable caching framework for fast generative inference. It calibrates the discrepancy between lightweight proxy signals and full-model computation, enabling efficient cache reuse while controlling approximation risk.