AI
DNeo: Your Parallel Draft Model Needs Better Target Feature Encoding
Publication: cs.AI updates on arXiv.org
Author: Jameson Sandler, Shane Bergsma, Mostafa Elhoushi, Bin Claire Zhang, Gavia Gray, Joel Hestness, Nolan Dey, Ferdinando Fioretto
Source published: 18/11/2026, 05:00:00 UTC
Briefing created: 18/11/2026, 05:53:26 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 18/11/2026, 05:03:24 UTC.
Briefings
A new speculative-diffusion architecture named DNeo combines a parallel drafter model with a larger target model to generate draft tokens in one forward pass, achieving 1.3x faster generation speed than the state-of-the-art DFlash-2 while maintaining high quality. The architecture introduces a working memory to encode previously rejected proposals and uses a mixture-of-features attention mechanism to condition the drafter on rich features computed by the target, addressing limitations in standard feature encoding such as lossy projections and memorylessness. This work was published on arXiv as a preprint and represents a significant advancement in high-quality, fast LLM inference.
Key points
- The DNeo architecture achieves 1.3x faster generation speed compared to the state-of-the-art DFlash-2 model.
- The DNeo model uses a working memory to encode previously rejected proposals, unlike standard memory-less approaches.
Context
Speculative-diffusion decoding uses a small parallel drafter model to generate draft tokens in one forward pass, which then undergo parallel verification via a larger target model. Faster inference can be achieved when multiple draft tokens pass verification simultaneously. Recent parallel draft architectures like DFlash and DSpark improve draft quality by conditioning on rich features computed by the target, through a mechanism known as feature encoding. However, standard feature encoding mechanisms have limitations including lossy projections, memorylessness, and uniform feature sampling.
Limitations
The evidence is a preprint from arXiv, v1, with no peer review or publication date. The study focuses on speculative-diffusion decoding and LLM inference; no specific city, region, or country is identified in the source. The architecture's performance and quality remain unverified beyond the claim of being state-of-the-art, and the specific impact on public applications is not described.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/11/2026, 05:53:26 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigymwslcnnzgbx4kmyj2mffvf67owpwshfuzlmjsorxbfeepw3isa