Another paper from us, this time on video token compression. Video VLMs have a token problem. A single video can produce tens of thousands of visual tokens, making prefill expensive and KV caches large. But aggressively pruning them risks deleting the details needed for temporal reasoning. CRAFT takes a different approach. It separates two decisions: Which tokens should merge? Global similarity decides. How should they merge? Position-aware weights and a channel-wise gate learn how to preserve the information. The process is recursive and query-agnostic, so the video can be compressed once and reused across multiple questions. At roughly 8× compression, CRAFT retains 96.8% of the backbone’s average accuracy while substantially reducing prefill and KV-cache costs. The takeaway: don’t just drop redundant tokens. Learn how to fuse them without losing the details.