The news. On July 8, 2026, the DeLS-Spec paper (arXiv:2607.07409) tackled a weakness in block-parallel speculative decoding: a drafter that predicts a whole block of tokens in one pass is fast, but it lacks explicit intra-block causal conditioning — each position is guessed without seeing the block's earlier positions. DeLS-Spec keeps the existing block drafter (DFlash) frozen as a long-context expert and adds a lightweight short-context local head, trained separately on next-token prediction; at inference the two drafters' logits are fused — a product of experts that also divides out a unigram prior — before the target verifies. On Qwen3 math, code, and dialogue benchmarks the authors report higher speedup and average acceptance length than DFlash. Read the paper →

Picture a worksheet of blanks and one helper who has read the whole page. She is fast: she glances at every blank and fills the entire row in a single pass. But because she writes them all at once, blank four is filled without reading what she just put in blank three — the words are each plausible on their own, yet they don't always flow into each other. Now hand the same row to a second helper who ignores the page and reads only the word just before each blank. Because he needs that preceding word filled in first, he has to work through the row in order rather than all at once — but each glance is quick, since he only ever looks at one word. He has no idea what the sentence is about, but he is excellent at the small question: given the last word, what comes next? DeLS-Spec is the trick of laying both helpers' answers over each blank and blending them — discounting the words that are common no matter what — so the whole-page helper keeps the sentence on topic and the last-word helper keeps it locally smooth. That blend is exactly what block-parallel speculative decoding was missing.

Underneath the metaphor, the bottleneck is how a block drafter guesses. Ordinary speculative decoding drafts one token at a time, each conditioned on the last; a block drafter like DFlash trades that away to go faster, predicting a whole block of positions in a single forward pass. The speed comes at a cost: the positions are computed in parallel, so a token near the end of the block is produced without conditioning on the drafter's own guesses for the tokens just before it. That missing link — intra-block causal conditioning — is why the later tokens in a block can read rougher, and rougher tokens are the ones the verification pass is likelier to throw away.

DeLS-Spec fixes the gap without retraining the block drafter. It freezes DFlash and re-labels it the long-context expert — the model that has seen the full prior context and is good at global coherence. Then it trains a separate, lightweight short-context local head with a plain next-token objective, so it becomes a short-context expert: it conditions only on a few immediately preceding tokens and is strong at exactly the local handoff the block drafter fumbles. This head runs sequentially over the drafted block, so it adds a little per-token work — but it is small enough that the extra accepted tokens more than pay for it, and the paper reports a higher end-to-end speedup than DFlash, not just higher acceptance. At inference, the two experts' logits are fused into a single distribution — a product of experts that adds a weighted short-context term to DFlash's logits and divides out a shared unigram prior so common tokens aren't counted twice — before the target model verifies the block. Every drafted token's score now blends the global signal from DFlash with the local signal from the small head, so the block tends to read smoother and, on the paper's math, code, and dialogue tests, the target keeps more of it — the draft variant changes the guess, not the target.

Where it earns its keep

Take a block of 6 drafted tokens (illustrative — the paper's real gains are on Qwen3 math, code, and dialogue, not these counts). Suppose the long-context block drafter nails the first few positions from the global context but, because positions 4–6 are guessed blind to their neighbors, they come out rough; the target model accepts through position 3 and rejects the rest, for an acceptance length of 3 out of the 6-token block. Now fuse in the short-context head: positions 4–6 finally get the given-the-last-word signal they were missing, so the wording locks and the target accepts through position 5 of the same block — an acceptance length of 5. The same DFlash block draft, now re-scored by the small local head and verified once, yields a longer accepted run — and because the head stays cheap while acceptance length rises, DeLS-Spec nets a higher end-to-end speedup than DFlash, not just higher acceptance. That is the effect DeLS-Spec reports across math, code, and dialogue.

Because the change lives in how the draft is scored rather than in the target model's weights, it should in principle stack with the rest of the toolkit — a better block drafter, a smarter verification rule, or a load-aware deployment are all orthogonal to it — and the real-world gain will depend on the model family and how predictable the text is. The headline is not a new drafter; it is a second, smaller one whose only job is the local handoff the first one skipped — fused in so a fast block draft picks up the local signal it lacked.

Goes deeper in: LLM Serving → Speculative Decoding → Draft Model Variants

Related explainers

• JetSpec — Parallel tree drafting — the other way to spend a draft budget better: branch the guess into a tree instead of fusing in a second scorer
• PPOW — window-level RL for speculative drafters — trains the drafter to miss less; DeLS-Spec instead fuses in a second head to miss less
• VIA-SD — Tiered confidence-gated verification — reworks the verify side of the same draft-verify loop
• EfficientRollout — Quantized self-drafters — shrinks the drafter itself rather than adding a second one