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

Interesting work, but I wonder whether KV-cache compression is solving the problem rather late in the pipeline. If an upstream evidence/state selection layer reduces the working context before prefill, the model never has to allocate KV-cache for most irrelevant tokens in the first place. In that architecture, the main question is not which tokens to evict after the full context entered the model, but which evidence should enter the context at all. So TIS seems most useful when long prefill is unavoidable or when no upstream epistemic compression exists. But for agentic or RAG-like systems, I would first compare it against upstream evidence/state compression, because that changes the problem from cache survival to context admission.