Triplet-Block Diffusion RWKV

Researchers introduced B^3D-RWKV, a diffusion-based RWKV variant designed for bidirectional discrete decoding.

Its triplet-block layout preserves RWKV’s O(L) inference efficiency while enabling parallel generation.

On an 8-task benchmark, the 7.2B model matched existing systems’ accuracy and decoded about 1.6x faster on average.

The work suggests a practical path to combine bidirectional diffusion with linear-time language model inference.