Reproduce “Thinking Outside the Box: Retention and Transmission of Information in Sliding-Window KV Inference” 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
Timothy DeLise, Seth Cromelin
Rolling-KV keeps a fixed 512-token cache while processing streaming text. Rather than rebuilding the final window from raw text, it carries forward previously computed states. In a synthetic four-choice secret-code task, Muse Glimmer 30B reaches 62.9% and Mistral 7B 39.9% in the hardest post-eviction region; other models remain near chance, limiting the finding to this task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.