Enabling KV Caching of Shared Prefix for Diffusion Language Models

In three linesDiffusion language models (DLMs) use bidirectional attention, invalidating standard KV caching techniques for shared prefixes. Researchers propose bicache, a method that dynamically identifies safe layer depth for reusing shared prefix KVs. Result: 36–98% throughput improvement without accuracy collapse.

Summary generated by Claude — human-verified