Presenting TIS (Token Importance Scoring) - A new way to compress KV cache

Your comment reminded me of some of the strategies employed by GLM5 detailed in this paper: https://arxiv.org/pdf/2602.15763