Reproduce “BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation” 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
Yaoyang Liu, Shuwei Shen, Jincheng Tang, Jinguang Liu, Jinbao Xue, Hailong Yang, Ke Zhang, Yinjun Wu
BELT compresses key and value caches for long-context question answering using low-rank representations designed to preserve attention behavior and allocating components across blocks under a fixed budget. Its key objective targets post-softmax attention distributions, while its value objective targets attention-output error. On Qwen3-32B at 70% compression, it averages 40.23 across seven LongBench tasks, versus STAR-KV’s 16.89; this is a benchmark average, not every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.