●builderWorth watching if you run long-context reasoning workloads where KV cache memory is a bottleneck — this could reduce memory without degrading multi-step reasoning quality.
●researcherForward Influence offers a new information-theoretic lens on token importance that challenges attention-only compression baselines.