Breaking01 / AI
B3 speeds up self-editing diffusion language models
The method needs no additional training or parameter tuning, and researchers reported generation quality on par with the base model.

Researchers reported that B3 delivered an 11.5–14.5% speedup for self-editing diffusion language models in a paper published on 11 Nov. The benchmark figures were published by the researchers.
According to the paper, B3 refines two blocks of text in parallel within a single model computation. It opens the next block only when the model is ready to begin generating it.
The paper describes B3 as a method that needs no additional training or parameter tuning. It applies to block diffusion models that revise their own generated text.
The researchers reported evaluating B3 across 10 datasets covering varied tasks and difficulty levels. They reported generation quality on par with the base model.
In the baseline approach described in the paper, revision of one block must finish before the next block can begin. Each block progresses only as fast as its slowest token, a small unit of text.
The paper's abstract did not disclose the hardware used to measure the speedup.