RCSD — Training speculative drafts from accepted and rejected sequences — What does it mean?
The news. On October 21, 2026, researchers listing Hisense contact addresses posted RCSD to arXiv. It trains EAGLE-style draft models with DPO or GRPO and reports a 35% longer average acceptance length than the HASS baseline and a 1.35× end-to-end speedup on ShareGPT, with generation quality unchanged. Read the paper →
Picture the sous-chef again. Each round they prep a handful of dishes, and the head chef tastes them in order and stops at the first dish that is wrong, so only the dishes accepted before the first rejection count. That is the draft-verify loop: the speedup is set by how many dishes survive before the first rejection, not by how closely the sous-chef copies the head chef's whole style.
The standard way to train the sous-chef ignores this. Soft-label distillation asks the draft to match the target's full probability distribution, which is like grading the sous-chef on every garnish the head chef might ever choose. A model with a small fraction of the target's parameters cannot match all of it, so some of its capacity goes to rare tokens. The paper argues that aligning the relative order of the target's likely top-k tokens can be enough for acceptance, so a draft can have a large distillation loss and still be accepted every time. The paper calls this a mismatch between the training objective (distribution matching, mode-covering) and the system goal (acceptance length, which rewards mode-seeking).
Token-by-token acceptance check
RCSD builds its training signals from target and draft sequences, with no human labels and no separate reward model. In the DPO version the chosen sequence is plain autoregressive output from the target. The rejected sequence is built by alternation: the target writes one token, the draft then writes a block of k tokens with no verification, and the two repeat. The target is the preference: the clean sequence should beat the one that contains unchecked draft blocks.
A masked loss keeps the signal clean. In the rejected sequence, the tokens that the target itself wrote are masked out, so the draft is not penalized for tokens that already agree with the target. The same mask is applied to the chosen sequence, because DPO is sensitive to sequence length. Only the draft's own stretches of the rejected sequence push the model.
The GRPO version replaces the pair with a group: the target-only sequence plus alternating sequences for several block sizes (the paper's example uses spec_step1 to spec_step3). Each sequence gets a reward made of two z-score normalized terms (each rescaled to mean 0 and spread 1 across the group), the target's log-probability of the sequence and the token accuracy, and the draft is updated with a clipped ratio and a KL penalty to a frozen copy of itself. For a deeper look at how the target decides, see the verification algorithm.
| Draft training | What the draft is rewarded for | Qwen3-1.7B, 6 drafted tokens: average acceptance length |
|---|---|---|
| HASS (baseline) | Matching the target's distribution | 2.26 (Table 1) |
| HASS + DPO | Beating a draft-contaminated sequence with the target-only one | 2.97 (Table 1) |
| HASS + GRPO | Group-relative reward from target log-probability and token accuracy | 3.00 (Table 1) |
Here is what those numbers mean in target-model work. Hold two things fixed: a 300-token reply, and the paper's Qwen3-1.7B setting with 6 drafted tokens per round, where acceptance length averages 2.26 for HASS and 3.00 with GRPO. Treating acceptance length as tokens committed per round and ignoring the cost of running the draft (illustrative), HASS needs about 300 / 2.26 ≈ 133 target passes and the GRPO-trained draft needs 300 / 3.00 = 100. Under those simplifying assumptions that is about 25% fewer verification rounds, a 1.33× cut in rounds; it is an estimate, not a measured pass count. The paper separately reports a 1.35× end-to-end speedup on ShareGPT. The paper also reports that the relative gain grows with longer drafts: on Qwen3-1.7B with GSM8K it rises from 31% at 3 drafted tokens to 49% at 6, which suggests a better-aligned draft pays off more the further it speculates.
Read the claim at its real scope. The experiments use two small models (Qwen3-1.7B and Llama-3.1-8B), one draft design (EAGLE), one baseline (HASS), and a draft trained on 90% of ShareGPT. The 35% is the abstract's headline figure; averaging the step-6 rows of Tables 1 and 2 myself gives about 33% for Qwen3-1.7B (2.26 to 3.00) and about 18% for Llama-3.1-8B (2.78 to 3.29), and the 1.35× speedup is for ShareGPT. The abstract describes pairs harvested from the accept/reject decisions of speculative inference, but in the method section the pairs are constructed by the scripted alternation above, and the paper gives no separate procedure for the continual-learning part of its name. Whether the same gains hold when pairs come from live serving traffic is not shown. For when the whole technique is worth the extra machinery, see when to use speculative decoding.
Goes deeper in: LLM Serving → Speculative Decoding → Draft Model Variants
Related explainers
- READY: speculative drafting from a recurrent state — another change on the draft side, this time to what the draft reads when it drafts.
- AMTP: acceptance survival over the draft horizon — how the draft horizon is chosen from how acceptance survives along the draft.
- EDR: expected decoding rounds with parallel drafters — a different training objective for drafters, tied to decoding rounds instead of per-token loss.