Cornell/IBM Method Lets Looped Transformers Share KV Cache, Cutting Memory by ~78%

A Cornell/IBM paper introduces a method (BeNice!) that enables looped transformers to losslessly share KV cache across iterations, cutting memory usage by 76-79% and addressing the growing cache problem in looped LMs.

2026-10-24 ~ 2026-10-24 · 2 related posts

• BeNice! method shares KV-cache across loops, scaling looped LLM depth without extra memory — stochasticchasm · 2026-10-24
• Cornell/IBM: looped transformers can share KV-cache across recursions, cutting memory 76-79% losslessly — WeijiaShi2 · 2026-10-24