Our paper AbridgeKV, on compressing long reasoning into the KV cache of learned pause tokens, is on arXiv!