D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding
arXiv cs.AI · Meng Wei, Yefei Chen, Antong Li, Yanan Gao, Xingliang Shi, Xin Gao, Zhiling Ye, Yun Yue, Chunxiao Guo, Peng Wei ·
arXiv:2611.10432v1 Announce Type: new Abstract: Speculative decoding efficiency depends on the number of consecutively accepted draft tokens, yet token-level training objectives do not directly optimize this quantity. We introduce D-REAL (Drafter Reinforcement of Expected Acceptance Length), a post-training method…
Excerpt provided by arXiv cs.AI via its public feed. Full reporting belongs to the original publisher.
Read the full story on arXiv cs.AI
Related coverage
The Verge AI · 1h ago
The indie game community is in disarray over AI and Big Walk
The Verge AI · 1h ago
Anthropic has its own AI-powered office suite now
The Verge AI · 4h ago
Adobe is going all in on agentic AI
TechCrunch AI · 5h ago
Cyberattacks threaten to make open-source AI a little less open, Hugging Face exec warns