PaperlayerView exact revision on arXiv

Source-linked AI summary

On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction

Eliezer Shlomi, Eliron Rahimi, Avi Mendelson, Chaim Baskin

arXiv:2610.22543v1cs.LG

TL;DR

DLMs face a speed–quality trade-off, and speculative decoding does not transfer directly because they must choose both what token to generate and where. SPOT predicts commit order from stale target distributions and verifies candidate trees with coupled noise; it reduces target passes and improves throughput over standard decoding while transferring across models.

  • Problem

    DLM decoding trades speed against quality, and speculative decoding must address both token choice and position choice because DLMs refine masked sequences rather than generate left to right.

    Source: §1 · pp. 1–2

  • Method

    SPOT uses a lightweight model to predict commit order from stale target distributions, builds candidate-state trees, and applies coupled-noise verification for sampling.

    Source: §1 · pp. 2–3

  • Results

    SPOT beats calibrated and training-free drafting baselines at T=0 and T=1, cuts target passes by up to 6.8×, and transfers across six DLMs without retraining.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The results establish commit-order prediction as a promising foundation for speculative decoding that accelerates DLM generation without sacrificing decoding quality or exactness.

    Source: §6 · pp. 9–10

    SPOT’s gains diminish when verification is costly or acceptance is low; verification becomes a bottleneck at wide blocks, and exact tree packing under bidirectional attention remains unresolved.

    Source: §6 · p. 10

Abstract

from arXiv · show

Diffusion language models (DLMs) promise parallel text generation, but in practice face an unresolved speed-quality trade-off. Exact decoding fills in only one masked position per model pass, while faster multi-token decoding schemes are lossy. Speculative decoding is the standard lossless acceleration for autoregressive models. However, it does not transfer directly to DLMs, which must choose both what to generate and where. We show that this order is the missing piece: the model's own predictions from the previous pass already give the right token at 92-100% of positions, but not which position comes next. Based on this observation, we introduce Speculating Positions Order Trajectory (SPOT), which learns to predict this order. A 0.62M-parameter model reads only the target's output distributions and proposes a tree of candidate continuations, and the target verifies all of them in a single pass by checking whether each equals its own next state. On the verification side, our analysis shows that prior block-diffusion verification approaches either assume $T=0$ or a sequential proposal interface. To support tree-based speculation at any temperature, we propose using coupled-noise match verification for diffusion blocks, following shared-Gumbel-noise coupling approaches. Our comprehensive experiments show that SPOT cuts target passes by up to 6.8x and reaches up to 3.2x the throughput of standard decoding, beating the calibrated and training-free drafting baselines in every task and budget at $T=0$ and $T=1$. In addition, we show that the learned drafter transfers across models, demonstrating successful transfer across six different DLMs. Project page: https://eliezer318.github.io/SPOT-dLLM/

1 INTRODUCTION

DLMs offer parallel generation but face a speed–quality trade-off, and speculative decoding does not transfer directly because DLMs choose both a token and its position. SPOT predicts commit order from stale target distributions and uses coupled-noise verification to support exact tree speculation under sampling.

1 INTRODUCTION

  • DLM decoding faces a speed–quality trade-off: conservative unmasking requires many denoising steps, while faster multi-position schemes can alter the trajectory and degrade quality.Unlike left-to-right AR decoding, block-diffusion models refine masked sequences with bidirectional denoising within blocks and causal generation across blocks.Source: §1 · pp. 1–2
  • In DLM speculation, committing one position changes the remaining masked positions’ representations and distributions, so the target must predict both what to write and where.The open questions are what a drafter should predict across several transitions and how to verify a speculative tree exactly under sampling.Source: §1 · p. 2
  • SPOT uses a 0.62M-parameter prefix-conditional drafter to predict commit order from stale output distributions and build multi-step trees of candidate block states.The target verifies drafted states in one batched forward and accepts the longest prefix consistent with its trajectory.Source: §1 · pp. 2–3
  • Coupled-noise verification shares per-position Gumbel randomness so target transitions are deterministic given the noise, enabling exact tree verification at arbitrary temperatures.Marginalizing over the shared noise recovers the target’s sampling distribution without requiring a sequential proposal interface.Source: §1 · pp. 2–3
  • SPOT cuts target passes by up to 6.8× and reaches up to 3.2× throughput over standard decoding, beating calibrated and training-free baselines at T=0 and T=1.The evaluation covers tasks, temperatures, block widths, and six DLMs; transferred checkpoints work across models without retraining.Source: §1 · p. 3

2 RELATED WORK

Prior speculative decoding methods for autoregressive models do not directly transfer to diffusion models, where bidirectional attention makes commit order part of decoding. SPOT combines learned order prediction, tree proposals, exact sampling verification, and cross-target transfer.

Speculative decoding and DLMs

  • Autoregressive speculative decoding verifies drafted tokens in parallel while preserving the target distribution, but bidirectional attention in DLMs makes the commit position an additional decision.Earlier work also developed auxiliary heads, learned drafters, dynamic draft trees, and diffusion drafters for autoregressive targets.Source: §2 · p. 3
  • SPOT carries shared-Gumbel coupling into diffusion decoding, making a target’s sampled commitment at a hypothetical block state deterministic given per-position shared noise.The related-work discussion describes this as extending communication-free coupling to diffusion decoding.Source: §2 · p. 3

Speculative decoding for diffusion targets

  • For DLM targets, SSD and Spiffy are training-free and greedy, MRP is greedy-only, BlockSpec uses trajectory trees, and SimSD verifies approximately rather than losslessly.SPOT instead learns commit order, builds multi-step trees, verifies exactly under sampling, and transfers across targets without retraining.Source: §2 · p. 3

3 ANALYSIS OF BLOCK-DIFFUSION SPECULATIVE DECODING

The analysis finds that commit-order information is more valuable for drafting than token identity, while verification cost limits realized speed. Under sampling, coupled and uncoupled matching preserve the target distribution, but coupling sustains acceptance as temperature rises.

3 ANALYSIS OF BLOCK-DIFFUSION SPECULATIVE DECODING

  • A draft is a hypothetical block state; the target verifies D drafts in one forward and accepts a draft only when it matches the target’s next commit at its parent.Because each draft costs a whole block, target forwards and wall-clock time can diverge.Source: §3 · pp. 3–4

3.1 DRAFT-TRAJECTORY DYNAMICS

  • Stale argmax tokens match committed tokens in 92% of first-block positions and at least 90% thereafter, while the previous round’s top-confidence position is next only 3% of the time by block nine.The stale top-eight candidates cover 89% of eight-commit continuations, and true positions outperform true tokens across tasks; the joint oracle reaches 3.8–4.6× peak throughput.Source: §3.1 · p. 4, Figure 2 · p. 4
  • SPOT predicts the next commit with 75% held-out accuracy versus 22% for stale-confidence ordering, recovering 54–81% of the joint wall-clock ceiling.Verification cost limits realized speed: the joint oracle’s 3.8–4.6× throughput ceiling is below its 8.0–9.7× NFE ceiling.Source: §3.1 · p. 4

3.2 VERIFICATION DYNAMICS ACROSS SAMPLING TEMPERATURES

  • Coupled and uncoupled matching preserve the target distribution, whereas greedy matching drifts as temperature rises; coupling keeps acceptance alive when samples depart from argmax.At T=2, uncoupled matching accepts almost nothing, while coupled matching still drafts; beyond about T=1.5, target quality becomes the dominant limitation.Source: §3.2 · p. 5

4 METHOD

SPOT treats commit-order prediction as the missing piece in diffusion-model speculation: it builds candidate state trees from stale target distributions and verifies them with shared-noise coupling. The verifier preserves the target trajectory while the drafter determines how many future transitions can be accepted.

4.1 SETUP

  • A DLM chooses both the token and masked position to commit; each commitment changes the remaining positions’ distributions, so speculative decoding must predict the commit order.At T=0 the target commits the highest-confidence position’s top token; at T>0 it samples from a temperature-adjusted conditional.Source: §4.1 · p. 6

4.2 THE DRAFTER: LEARNED COMMIT ORDER

  • A 0.62M-parameter transformer predicts the next committed position from stale target conditionals and hypothetical prefixes, then pairs each position with stale-distribution token hypotheses to build a tree.It uses no target hidden states, token embeddings, or text; training uses draft-free trajectories and prefix-conditioned target commitments as supervision.Source: §4.2 · p. 6

4.3 DIFFUSION-BLOCK VERIFICATION BY COUPLED-NOISE MATCH

  • At T=0, state-equality verification accepts a drafted edge only when it matches the target’s deterministic next state; at T>0, shared Gumbel noise makes the transitions comparable.For any drafter, the verifier preserves the target trajectory or distribution; draft quality determines how many future transitions are accepted in one target forward.Source: §4.3 · p. 7
  • Coupled verification is more effective when stale and fresh sampling distributions are closer, with match probability 1 + δT, where δT is their total variation distance.Rejection sampling can also verify exact proposals, but classical per-link rejection sampling assumes a sequential proposal interface, whereas coupled-noise match works directly with batched draft trees.Source: §4.3 · p. 7

5 EXPERIMENTS

SPOT outperforms the evaluated drafting baselines across tasks and budgets, including at T=1. Experiments also test transfer across targets and isolate the contributions of the tree, token hedge, mass floor, and padding.

5.2 MAIN RESULTS

  • At D8/D20, SPOT uses 3.2–4.7×/4.1–6.8× fewer target forwards than draft-free decoding; versus Spiffy it has 15–35% fewer NFE and 1.1–1.5× throughput in every task-and-budget setting.At D8 it reaches 2.0–2.7× single-device wall-clock speed versus draft-free decoding.Source: Table 2 · p. 8, §5.2 · p. 8

5.1 EXPERIMENTAL SETUP

  • The experiments cover GSM8K, MBPP, HumanEval, and MATH500 on six DLMs, with budgets D ∈ {8, 12, 20}; baselines share the decoder and verifier to compare drafting alone.The protocol reports NFE per sample, accuracy, and throughput relative to a draft-free run on the same task, model, and temperature.Source: §5.1 · p. 8

5.3 SAMPLING TEMPERATURE: COUPLED-NOISE MATCH AGAINST THE BASELINES

  • At T=1, SPOT has 15–34% fewer NFE than Spiffy and 1.15–1.48× its throughput; results are distribution-preserving, with accuracies within 0.06 of the control over seeds.The free chain is the strongest baseline on MBPP NFE but trails SPOT in throughput by 0.47–0.92, and reaches 116 NFE and 0.60× at D20 on MATH500.Source: §5.3 · pp. 8–9

5.4 TRANSFERABILITY WITHIN AND ACROSS MODEL FAMILIES

  • A checkpoint transferred across targets without retraining or recalibration; transferred runs cost 3–7% more NFE on SDAR-8B and 3–11% on LLaDA2.0-mini, while beating every SDAR-8B-calibrated baseline on the former and every baseline on GSM8K and MATH500 on the latter.For LLaDA-8B and Dream-7B, which use an approximate cache and are not distribution-preserving, transferred checkpoints reduce 256 forwards to 55–88 and 51–79, respectively.Source: §5.4 · p. 9, Table 3 · p. 9

5.5 ABLATIONS

  • Replacing the tree with a chain increases NFE by 7.3 at D8 and 11.9 at D20, while throughput ratio falls from 1.72 to 1.34 on SDAR-8B MBPP.The token hedge and scaled corpus each change NFE by under one; the mass floor trades NFE for wall-clock speed only without padding.Source: §5.5 · p. 9

6 DISCUSSION

The discussion presents commit-order prediction as the missing signal for effective DLM speculation and reports faster generation without sacrificing decoding quality or exactness. It also identifies settings where SPOT’s gains diminish.

6 DISCUSSION

  • SPOT combines stale-distribution token hypotheses, a 0.62M-parameter commit-order drafter, and coupled-noise verification to reduce target forwards and transfer across six DLMs without retraining.The paper reports 15–35% fewer target forwards than an offline-calibrated draft graph, at up to 1.5× its throughput.Source: §6 · p. 9
  • The authors conclude that faster DLM generation can preserve the quality and exactness of the underlying decoding process, with commit-order prediction a promising foundation for speculative decoding.The stated scope is the paper’s reported results; the discussion does not claim that all settings benefit equally.Source: §6 · p. 10
  • SPOT’s gains diminish when verification is costly or acceptance is low; at T ≥ 2, the T=0-trained order model is outperformed by noise-aware ranking.The paper also notes that exact tree packing remains unresolved under bidirectional attention.Source: §6 · p. 10

AI USE STATEMENT

This material states the paper’s AI-use and reproducibility disclosures, then formalizes decoding and proves that state-equality verification preserves the draft-free greedy trajectory.

AI USE STATEMENT

  • The authors report using generative AI for language editing and experimental-code implementation, but not for formulating questions, developing the method, defining evaluation, or drawing conclusions.They state that authors inspected and verified AI-assisted contributions and remain responsible for the work.Source: AI USE STATEMENT · p. 10
  • The work concerns inference acceleration in DLMs, reports no human subjects or personal-data collection, and says benchmark numbers come from one run on one machine.The machine had eight identical NVIDIA RTX 6000 Ada cards, with one job per card.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 FORMAL DECODING MODEL

  • The formal setup assumes fixed causal context during an active block, deterministic tie-breaking for position selection, and target outputs invariant to physical batch shape.These assumptions separate algorithmic exactness from finite-precision effects of batch shape.Source: §A.1 · p. 14

A.2 PROOF OF PROPOSITION 1

  • At T=0, state-equality verification preserves the complete draft-free generated sequence for any finite draft tree.The accepted states form a single path because the target has a unique successor at each state.Source: §A.2 · p. 16

A.3 PROOF OF PROPOSITION 2

The proof constructs temperature-T sampling from shared Gumbel noise and applies deterministic state-equality verification conditional on that noise. It also introduces rejection sampling as an exact alternative when proposals are sequential.

A.3 PROOF OF PROPOSITION 2

  • Independent Gumbel rows at previously uncommitted positions make the Gumbel-driven decoder follow the ordinary temperature-T target token law along its adaptive trajectory.The selected position depends on the current state, not on any unused Gumbel row.Source: §A.3 · pp. 16–17
  • For each fixed shared Gumbel realization, state-equality verification reproduces the deterministic target trajectory; averaging over the noise preserves the draft-free trajectory distribution.The argument permits the drafter to depend on stale target distributions, learned parameters, decoding history, or the shared randomness.Source: §A.3 · p. 17

B.1 SEQUENTIAL PROPOSAL SETUP

  • Rejection sampling is presented as an exact alternative when the drafter supplies an explicit normalized sequential proposal distribution over position-token commitments.The paper states that its proof requires proposals to be sampled from the distribution supplied to the verifier; the described chain proposal factors into a next-position distribution and stale token conditional.Source: §B.1 · p. 18

B.2 EXACTNESS AND OPTIMAL ACCEPTANCE

The section proves that accept–residual rejection sampling is exact and achieves maximal per-link agreement, then distinguishes its sequential interface from batched-tree verification and reports implementation checks.

B.2 EXACTNESS AND OPTIMAL ACCEPTANCE

  • Pr[accept | yk] = 1 − DTV(πk, qk) is the maximal agreement probability, and sequential use reproduces the complete draft-free target distribution.The accept–residual rule therefore realizes a maximal coupling of the target and proposal distributions.Source: §B.2 · pp. 18–20

B.3 RELATION TO BATCHED TREE VERIFICATION

  • Classical per-link rejection sampling does not directly verify a preconstructed tree with multiple candidate children at one predecessor or exploit its full batched structure.After rejection, the residual sample need not belong to the preconstructed tree.Source: §B.3 · p. 20

B.4 EMPIRICAL VALIDATION

  • The implementation’s measured rejection-sampling acceptance rate differed from the theoretical value by at most 0.002 across tested states.The checks also compared empirical committed position-token distributions with target distributions, including deliberately miscalibrated proposals.Source: §B.4 · p. 20

C DRAFTER TRAINING AND DRAFT-TREE CONSTRUCTION

The drafter is trained on target-generated trajectories using stale conditionals and replayed hypothetical prefixes. At inference, it predicts commit positions and combines those predictions with stale token hypotheses to construct a budgeted tree.

C.1 TRAINING THE COMMIT-ORDER PREDICTOR

  • Training examples replay prefixes of draft-free target trajectories while keeping the original stale conditionals fixed, so supervision matches the information available during recursive inference.The next target commitment after each replayed prefix is the supervision target.Source: §C.1 · p. 21

C.2 PREFIX-CONDITIONAL DRAFT-TREE CONSTRUCTION

  • At each hypothetical state, the order model predicts the next committed position, which is paired with token hypotheses from the stale target conditionals to form candidate extensions.At T=0, the stale runner-up is retained when top-1 probability is below 0.6; at T>0, token hypotheses use the coupled stale draw.Source: §C.2 · p. 21
  • DDTree retains the highest-mass candidate states within the draft budget, using prefix mass formed from position probabilities and stale token probabilities.Because child mass cannot exceed parent mass, level-wise expansion recovers the same top-D fixpoint as sequential best-first search; the final model uses calibration temperature τ = 1.Source: §C.2 · p. 22

D.1 DRAFTER TRAINING DETAILS

Across the covered sections, SPOT trains a small order predictor on target-generated trajectories and verifies speculative states through isolated block replicas. Its performance varies with target and budget economics, while transfer and longer generations retain substantial gains.

D.1 DRAFTER TRAINING DETAILS

  • The 0.62M-parameter drafter is a three-layer transformer trained on target T=0 trajectories, with held-out next-commit accuracy of 0.75 for the released SDAR-8B checkpoint.It uses stale top-32 probabilities and path features; no evaluation samples appear in the training pools, but SDAR-8B and mini training uses code only, leaving their math-task runs out of domain.Source: §D.1 · p. 22

D.2 DRAFT LAYOUT AND VERIFICATION COST

  • A verification forward evaluates the root and each drafted state as separate block replicas, so each replica’s logits match the target at that state and rejection does not change output.Drafted forwards process (1+D)L positions, which makes target-forward counts and wall-clock throughput diverge; the benchmark uses exact prefix caching where available and shared draft-free controls.Source: §D.2 · p. 23, §D.8 · p. 25

D.5 MASS-FLOOR CALIBRATION

  • The mass-floor setting is calibrated per target on 50 held-out MBPP samples; on SDAR-8B, γ=0.1 leads γ=0.05 by 10% in mean throughput ratio.On LLaDA2.0-mini no floor beats γ=0, so the floor is disabled there.Source: §D.5 · p. 23, Table 4 · p. 24

D.6 METHODS OUTSIDE THE COMPARISON

  • The paper excludes SimSD because it is lossy and MRP because its results use different implementations, comparing only relative gains for MRP.MRP’s reported comparison favors SPOT in all twelve settings, but its argmax-equality test has no distribution-preserving sampling extension.Source: §D.6 · pp. 23–24, Table 5 · p. 24

D.7 FULL-ATTENTION TARGETS (LLADA-8B, DREAM-7B)

  • The full-attention LLaDA-8B and Dream-7B transfer rows use an approximate prefix cache and are reported against cached controls, not as lossless claims.The replica verification is exact relative to the cached decoder; the approximation lies in the cache.Source: §D.7 · p. 24

F.3 MORE MODELS

  • The same recipe transfers across targets: SPOT leads every SDAR-4B comparison, while on LLaDA2.0-mini it has lowest NFE in all 36 and higher throughput in 34.On SDAR-1.7B, padding costs can let Spiffy win throughput columns despite SPOT using 14–23% fewer forwards; SPOT-γ recovers all but one at T=0 and all at T=1.Source: Table 8 · p. 26, §F.3 · p. 26, Table 9 · p. 27

F.4 MORE BLOCK SIZES

  • At block width 16, the joint oracle reaches 6.2–7.2× NFE and 5.1–5.7× honest wall-clock ceilings; a frozen recipe matches or beats separately tuned drafters across four settings.The frozen drafter’s reported advantages range from 0.4–1.0 NFE on SDAR b32 to 4.6–6.5 NFE on mini b16.Source: §F.4 · p. 29, Table 15 · p. 29

F.5 GENERATION LENGTH

  • At length 512, SPOT-γ reaches 3.23× draft-free throughput on MBPP at T=1, while SPOT is ahead of every baseline in NFE and throughput at both budgets.At length 64, padded SPOT at D20 falls below draft-free decoding (0.89×); the unpadded SPOT-γ is flat at 1.87× for D8 and D20.Source: §F.5 · p. 29, Table 16 · p. 29

F.6 MORE ABLATIONS AND CURVES

The appendix examines how budget, temperature, verifier choice, and block depth affect SPOT and its baselines. It finds that throughput-optimal budgets and rejection causes depend on verification costs and the depth of correctly ordered paths.

F.6 MORE ABLATIONS AND CURVES

  • At D20, SPOT’s throughput falls to 1.25–1.87× on SDAR-8B tasks, while every task’s padded SPOT throughput optimum is D6.NFE remains monotone through D20; cached mini instead rises slowly to a D12–D20 plateau, with 2.22× at D16.Source: §F.6 · p. 31, Table 20 · p. 31, Figure 3 · p. 31
  • At T=2, uncoupled acceptance nearly disappears, while under coupling Spiffy’s noise-aware ranking uses 114.6 NFE at D8 versus 187.4 for the T=0-trained order model.At T=1.5 and D20, coupled versus uncoupled SPOT uses 54.1 versus 62.7 NFE; the near-uniform T=2 setting is described as producing unusable text.Source: §F.6 · p. 31, Table 21 · p. 32

F.7 ADDITIONAL VERIFIER MATERIAL

  • At T=1, SPOT’s tree saves 12–17 NFE over its chain under the same verifier; on chains, match and rejection sampling differ by at most 3 NFE.For SPOT’s chains, match has a 0.09 higher throughput ratio at D8, while rejection sampling’s 3.0 fewer forwards offset its test cost at D20.Source: §F.7 · p. 32, Table 22 · p. 32

F.8 MECHANISM: WHAT HAPPENS IN DEEP BLOCKS

  • The previous round’s top-confidence position still commits with probability only 0.03 by block 8, although stale argmax tokens remain correct at 0.92–1.00 across blocks.The learned prefix-conditional order model reaches 14.5 accepted tokens per round, versus about 5 for the fixed graph and 3.4 for the free chain.Source: §F.8 · p. 33, Figure 4 · p. 33

F.9 ERROR ANALYSIS OF THE DRAFTER

  • Position misses explain 62–81% of first rejections, while at D20 tree-budget exhaustion accounts for 32% on MBPP and 90% in the last block.The prefix-mass score is nearly calibrated: acceptance is 96% in the top mass bin and 3% in the bottom.Source: §F.9 · p. 33, Figure 5 · p. 34

F.10 DRAFTER ARCHITECTURE

  • An offline capacity probe found that 8× the parameters improved held-out next-commit accuracy by only 0.004; the released drafter size was not tuned on the final benchmark.The probe instead found that broadening training data moved decoding.Source: §F.10 · pp. 33–34

G REPRODUCIBILITY

The appendix records reproducibility details and defines terms used to describe SPOT’s drafting, verification, and performance measures. It also specifies the method’s adaptive budget and coupled-noise verification terminology.

G REPRODUCIBILITY

  • Code, trained drafters, and reproduction instructions are released under the MIT license, with drafters also mirrored on Hugging Face.The appendix identifies the released resources and their locations.Source: G REPRODUCIBILITY · p. 34

H.1 TERMS

  • Coupled-noise match shares Gumbel noise between drafter and target so the target transition is deterministic conditional on that noise and verification reduces to state equality.NFE counts target forwards; throughput ratio is normalized against the draft-free run for the same task, model, and temperature.Source: §H.1 · p. 35
  • SPOT’s draft tree recursively grows candidate block states from stale target distributions, prioritizing high-probability branches up to a budget of D states.SPOT-γ discards candidates below prefix-mass floor γ, adapting the realized verification budget to drafter confidence.Source: §H.1 · p. 35

H.2 COMMON QUESTIONS

SPOT addresses the difficulty of speculative decoding in DLMs, where the next position as well as token must be predicted. It learns commit order and verifies candidate continuations, reporting gains across models alongside boundaries on where those gains hold.

H.2 COMMON QUESTIONS

  • DLM speculation must predict both what token to write and which masked position to commit next, unlike autoregressive speculation’s fixed next position.The stale argmax token is correct in 92% of first-block commits and at least 90% thereafter, while the previous round’s top-confidence position commits next only 3% of the time by the ninth block.Source: §H.2 · p. 35
  • SPOT predicts commit order from stale top-32 probabilities and grows a tree of hypothetical states, which the target checks in one batched forward.The drafter trains on draft-free target trajectories without target hidden states, token embeddings, or text.Source: §H.2 · p. 35
  • At T>0, shared per-position Gumbel noise makes the target transition deterministic conditional on noise, preserving the draft-free output distribution for targets with an exact prefix cache.The paper distinguishes this exact regime from approximate-cache transfer rows.Source: §H.2 · pp. 35–36
  • Across six DLMs, SPOT’s recipe transfers: on LLaDA2.0-mini it has the lowest NFE in all 36 comparisons and higher throughput in 34.For SDAR-1.7B, Spiffy wins some throughput columns; LLaDA-8B and Dream-7B transfer rows use an approximate prefix cache and demonstrate reach rather than losslessness.Source: §H.2 · p. 36
  • Gains diminish when verification is costly or acceptance is low; at T≥2, noise-aware ranking outperforms the order model trained at T=0.The paper also identifies unresolved exact tree packing under bidirectional attention.Source: §H.2 · p. 36

Related papers

Loading 2610.22543v1…