Reproduce “AbridgeKV: Compressing Long Reasoning into the KV Cache of Learned Pause Tokens” 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
Jack Cai, Lijie Yang, Vasilis Kontonis, Dimitris Papailiopoulos, Shivam Garg
ABRIDGEKV compresses long mathematical reasoning using learned pause tokens to replace completed 512-token blocks’ key/value cache entries with compact entries for later tokens, while leaving visible reasoning intact. On 446 competition problems, it scored 43.0 avg@8 versus 43.8 for full KV, a statistically compatible result, using 369 versus 1,277 MiB per sequence in the tested Qwen3-4B-Base setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.