SinkRouter: Sink-Aware Routing for Efficient Long-Context Decoding in Large Language and Multimodal Models

arXiv:2604.16883v1 Announce Type: cross Abstract: In long-context decoding for LLMs and LMMs, attention becomes increasingly memory-bound because each decoding step must load a large amount of KV-cache data from GPU memory. Existing acceleration strategies often trade efficiency for accuracy by relying on heuristic pruning that may discard useful information. At a deeper level, they also tend to indiscriminately preserve all high-scoring tokens, treat early tokens as indispensable anchors, or re