Last released Aug 20, 2026
Adaptive KV-cache compression for LLM inference via dynamic token eviction and bit allocation.