Speculative decoding usually lets a small model draft and a large model verify in parallel. Block diffusion pushes this further: a lightweight diffusion drafter emits an entire block of tokens in one forward pass instead of guessing one by one. But the D-Loop paper, released on October 5 by City University of Hong Kong, The Chinese University of Hong Kong and Huawei Research, identifies a systematic flaw they call the "repetition trap": every position inside the block predicts its own marginal distribution without seeing its neighbors, so adjacent positions often emit the same token over and over. Once a draft repeats, verification cuts it short, the acceptance length stalls, and the speedup shrinks.
How bad is the repetition trap
The paper measures repetition in raw DFlash drafts (acceptance lengths on Qwen3-8B, sampling). The adjacent-repeat rate is 17.4% on GSM8K, 29.4% on LiveCodeBench, and 42.6% on MT-Bench; consecutive repeated content words appear in 72.5%–94.5% of draft blocks, and the longest repeated run on MT-Bench averages 4.7 tokens. The corresponding acceptance lengths are 5.28, 3.26 and 3.10 — heavier repetition, shorter acceptance.
Two looped passes, one model
Earlier fixes bolt a causal head onto the drafter or train a separate drafter, adding parameters and training objectives. D-Loop does the opposite: no new modules, just the same drafting backbone run twice. The first pass proposes a full block; the second conditions on a selected prefix and regenerates the suffix in parallel. A complementary prefix–suffix objective trains the shared parameters for both anchor-only prefix prediction and prefix-conditioned suffix prediction. With only the second loop (K=1), adjacent-repeat rates drop by 73.7%, 83.0% and 79.8% on GSM8K, LiveCodeBench and MT-Bench.
The speedup numbers
Across eight math, code and chat benchmarks (including GSM8K, MBPP, LiveCodeBench, MT-Bench and Alpaca), D-Loop achieves the longest acceptance length among the methods the authors ran themselves on Qwen3-4B at both temperatures, with average speedups of 5.06x (greedy) and 4.11x (sampling); on Qwen3-8B it reaches average acceptance lengths of 7.19 and 5.77 with average speedups of 5.02x and 4.03x. The sampling regime matters most: the average-speedup edge over DSpark widens from 2.2% to 15.4% on the 4B model and from 9.7% to 26.2% on the 8B — higher temperature worsens the marginal-distribution repetition disease, which makes the second loop more valuable.
So what
For deployment teams, sampling is the everyday mode for chat and creative generation, and that is exactly where D-Loop's gains grow — with no extra parameters and no second training objective, which keeps integration cheap. Caveats: DFlash and DSpark comparisons come from the authors' own runs, and EAGLE-3 numbers are quoted from the DFlash paper (tree size 60), so other tree sizes remain open for discussion. Still, the route of "trading loops for quality and shared parameters for deployment cost" adds a key piece to diffusion-style speculative decoding — see arXiv:2610.06011.