DNeo Improves Parallel Draft Model Feature Encoding
November 15, 2026
Standard feature encoding in speculative-diffusion decoding suffers from lossy projections, memory-less processing, and sampled feature limitations. DNeo addresses these by improving how parallel drafter models condition on target model features to boost inference speed.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to fix residual mixing and memory issues in speculative decoding architectures.