Metacognitive Decoding: Efficient Inference for Diffusion Language Agents
A controller decides how much text a diffusion language agent must generate for each step.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv