AI RESEARCH
LongFlow: Efficient KV Cache Compression for Reasoning M
arXiv CS.LG
•
ArXi:2603.11504v1 Announce Type: new Recent reasoning models such as OpenAI-o1 and DeepSeek-R1 have shown strong performance on complex tasks including mathematical reasoning and code generation. However, this performance gain comes with substantially longer output sequences, leading to significantly increased deployment costs. In particular, long outputs require large KV caches, resulting in high memory consumption and severe bandwidth pressure during attention computation.