Reproduce “ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zican Dong, Peiyu Liu, Junyi Li, Zhipeng Chen, Han Peng, Shuo Wang, Wayne Xin Zhao
ForesightKV learns to preserve key–value entries likely to matter later, rather than relying only on recent attention. It trains a lightweight scorer with future-attention supervision and refines eviction through reinforcement learning. On AIME2024, Qwen3-4B with a 1K cache scored 54.5 pass@1, versus 44.8 for R-KV at 2K. This benchmark comparison does not establish gains across tasks or models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.