LLM Digest

Story

arxiv_cs_ai · Nov 7, 2026 · paper

Source brief

D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding

arxiv.orgNov 7, 2026
original source linked

In brief

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 E...

Feed lens

eval

Read the original at arxiv.org →Open in live feed Read that day’s brief

D-REAL: Acceptance-Length Optimization with Unbiased Gradients for Speculative Decoding

Earlier in this thread 4 items

Training Parallel Speculative Draft Models by Directly Minimizing Expected Decoding Rounds

AsymSpec: Context-Asymmetric Speculative Decoding for Agentic LLMs

Give the Drafter a Second Chance to Guess: Parallel Reconsideration for Speculative Decoding

NgramRL: Accelerating RL Post-Training for Structured Generation via Rollout-Aware N-gram Speculative Decoding