EDA adapts draft models to fine-tuned LLMs via lightweight private components, regenerated training data, and selective sampling, restoring speculative decoding performance at much lower cost than full retraining.
HDR integrates hierarchical tree-structured latents into causal video generation to enable coarse-to-fine multi-step visual reasoning with sparse attention, boosting reasoning success by 76% over streaming diffusion while running 54x faster than bidirectional diffusion.