PaperlayerView exact revision on arXiv

Source-linked AI summary

Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules

Amr Mohamed, Yang Zhang, Michalis Vazirgiannis, Guokan Shang

arXiv:2512.02892v1cs.CL

TL;DR

Diffusion language models decode iteratively, making sampling efficiency a central bottleneck. SchED uses a training-free, progress-aware confidence schedule to stop decoding when predictions stabilize. Across Dream and LLaDA variants, it delivers substantial speedups with near-baseline quality, particularly for instruction-tuned models.

  • Problem

    Diffusion language model decoding requires many refinement steps, while preset step budgets and transfer schedules can waste computation or destabilize quality.

    Source: §1 · p. 1

  • Method

    SchED aggregates full-span logit margins and exits when they meet a smooth, progress-dependent confidence threshold, without retraining the model.

    Source: §1 · p. 2, §3.2 · p. 5

  • Results

    Instruction-tuned models retain 99.8–100% of baseline performance with 3.8–4× speedups, while base models retain 99.1–100% with 1.04–1.14× conservative speedups and up to 2.34× under aggressive schedules.

    Source: §1 · p. 2

  • Takeaways & Limitations

    SchED reduces denoising steps across Dream and LLaDA, with particularly large gains on instruction-tuned variants where uncertainty decays rapidly.

    Source: §8 · p. 9

    The right schedule is task-dependent, and the paper leaves learned aggregation, joint training, and tighter integration with complementary acceleration techniques for future work.

    Source: LIMITATIONS · p. 10

Abstract

from arXiv · show

Diffusion large language models (dLLMs) offer a promising alternative to autoregressive models, but their practical utility is severely hampered by slow, iterative sampling. We present SchED, a training-free, model-agnostic early-exit algorithm that aggregates full-span logit margins and halts decoding once a smooth, progress-dependent confidence threshold is met. We evaluated SchED on two dLLM families (Dream and LLaDA), in base and instruction-tuned variants across ten benchmarks spanning downstream tasks including multiple-choice question answering (MCQ), math, long-form QA/summarization, and translation. SchED delivers large, stable accelerations: on instruction-tuned models, it achieves $3.8$-$4.0\times$ speedups while retaining $99.8$-$100\%$ of the baseline score on average. On base models, SchED yields consistent speedup gains with $99.1$-$100\%$ performance retention, with up to $2.34\times$ under more aggressive settings. Using a conservative speed metric that heavily penalizes quality loss (QPS, $\gamma{=}4$), we show that SchED is robust and clearly outperforms prior confidence-based early-exit methods, which break down on long-form generation. An entropy analysis of the model's token predictions reveals that instruction tuning speeds up the decay of predictive entropy. By turning genuine confidence stabilization into computational savings, SchED makes dLLM decoding substantially more efficient.

1 INTRODUCTION

Diffusion language models offer parallel decoding but remain costly because iterative refinement and conservative step budgets waste computation. SchED uses a smooth, progress-dependent confidence threshold to stop when predictions stabilize, achieving large speedups with near-baseline quality, especially on instruction-tuned models.

1 INTRODUCTION

  • dLLMs offer parallel refinement, flexible infilling, and bidirectional attention, but their iterative reverse-diffusion steps and conservative preset budgets make decoding inefficient.Extreme heuristic step budgets can also produce unstable behavior across tasks.Source: §1 · p. 1
  • Prior acceleration approaches include dynamic policies, error-correction methods, few-step generation, caching, and speculative decoding, but can require training, auxiliary models, or complex heuristics.These approaches motivate a simpler training-free, architecture-agnostic early-exit principle.Source: §1 · p. 2
  • SchED aggregates full-span top-2 logit margins and compares them with a smooth, progress-dependent threshold, enabling training-free, architecture-agnostic stopping without fixed budgets.It composes with standard transfer schedules, including single-suffix and block diffusion.Source: §1 · p. 2
  • Instruction-tuned models retain 99.8–100% of baseline performance with 3.8–4× speedups, while base models retain 99.1–100% with 1.04–1.14× conservative speedups and up to 2.34× under aggressive schedules.SchED also outperforms prior early-commit methods on a conservative quality-penalized speed metric.Source: §1 · p. 2
  • Predictive-entropy analysis shows that instruction tuning accelerates confidence stabilization, aligning with progress-aware thresholds and explaining early exits without quality loss.The analysis tracks entropy trajectories over the answer span.Source: §1 · p. 2

2 RELATED WORK

Prior dLLM work establishes the model family and pursues efficiency through early commitment, dynamic policies, few-step generation, caching, and speculation. SchED treats efficiency as a stopping decision, using a smooth confidence schedule rather than a fixed or abrupt rule.

Diffusion LLMs

  • dLLM research includes training from scratch and adapting autoregressive checkpoints, alongside advances in sparse mixture-of-experts, long context, high-throughput inference, and scaling analyses.The related work describes LLaDA and Dream base/instruct variants among these developments.Source: §2 · p. 3

Sampling efficiency and early commit

  • SchED differs from Prophet by smoothly relaxing confidence thresholds as generation progresses, enabling stable, training-free early exit across architectures.Prophet uses top-2 logit gaps for training-free early commitment.Source: §2 · p. 3
  • Efficiency efforts span early-commit and dynamic policies, few-step or one-step generation, caching that reduces per-step latency, and speculative draft-and-verify decoding.These approaches target either step counts or latency within each step.Source: §2 · p. 3

3 METHODS

The methods section formalizes discrete masked diffusion and introduces SchED as an inference-time stopping rule. SchED aggregates confidence over the answer region and exits when it exceeds a smooth threshold that relaxes with diffusion progress.

3 METHODS

  • The methods section formalizes discrete diffusion language models before presenting confidence aggregation, progress-dependent schedules, and the early-exit algorithm.Source: §3 · p. 3

3.1 PRELIMINARIES: DISCRETE DIFFUSION LANGUAGE MODELS

  • Discrete diffusion progressively masks clean sequences, then models a reverse chain that unmasks tokens using per-position categorical predictions.Transfer schedules select masked positions for updates; Dream may update a suffix or proportion, while LLaDA may select contiguous blocks.Source: §3.1 · p. 4
  • The standard training objective teaches predictions of original tokens at masked sites from partially noised contexts, using uniformly sampled timesteps.The section defines the answer region and normalized progress used later in confidence aggregation and thresholding.Source: §3.1 · pp. 3–4

3.2 SchED: SCHEDULE-BASED EARLY DECODING

  • Confidence is measured by aggregating top-2 logit margins over the full answer region, using current-step logits for each position.The thresholds are expressed in logit units.Source: §3.2 · p. 5
  • SchED exits when aggregated answer-region confidence meets a smooth, nonincreasing threshold indexed by normalized diffusion progress, rather than waiting for a fixed step budget.The complete procedure uses the existing model and transfer schedule, filling remaining masked tokens with current argmax predictions when it exits.Source: §3.2 · p. 5
  • The threshold schedule is parameterized by initial and final confidence thresholds, with an optional slope; its linear and exponential forms provide smooth trajectories between the endpoints.Different schedules offer different degrees of early-exit aggressiveness and stability control.Source: §3.2 · p. 5

4 EXPERIMENTAL SETTINGS

The evaluation tests SchED across model families, task types, and quality measures, comparing it with standard diffusion sampling and Prophet under multiple smooth-threshold settings.

4 EXPERIMENTAL SETTINGS

  • Against standard diffusion sampling and Prophet, each schedule family is tested with τhigh=7.5 and τlow=0 or 2.5.These lower-threshold settings produce relaxed and more conservative regimes, respectively.Source: §4 · p. 6
  • Dream and LLaDA models are evaluated in base and instruction-tuned variants, with low-confidence remasking applied across all models.Dream uses full-sequence refinement, while LLaDA denoises contiguous token segments.Source: §4 · p. 6
  • Ten benchmarks cover multiple-choice, math, long-context QA and summarization, and translation tasks.The suite includes GPQA, GSM8K, HellaSwag, MMLU, PIQA, Winogrande, HotpotQA, MultiNews, WMT14 En–Fr, and WMT16 En–De.Source: §4 · p. 6
  • MCQ and GSM8K use accuracy, HotpotQA uses token-level F1, MultiNews uses ROUGE, and translation uses CHRF.Evaluations use the Language Model Evaluation Harness.Source: §4 · p. 6
  • QPS jointly captures quality and speed, with γ controlling the penalty for quality degradation; experiments use γ=4 for conservative comparisons.Higher γ values emphasize fidelity over raw acceleration.Source: §4 · p. 6

5 RESULTS

SchED produces substantial speedups at near-parity quality on Dream Instruct, with more modest conservative gains on Dream Base; LLaDA shows similar trends, and QPS comparisons favor SchED over Prophet.

5 RESULTS

  • Dream Base conservative schedules yield 1.04–1.14× average speedups near parity; setting τlow=0 raises speed to 2.34× with an ≈2% overall quality cost.The 2.34× setting shows its largest task-specific gain on HotpotQA (∆+7.1).Source: Table 1 · p. 6, §5 · p. 7
  • On Dream Instruct, conservative smooth schedules deliver ≈3.8–4.0× average speedups near parity, while Exp-k=16 reaches 4.48× with a 57.59 average score.The cited table reports Dream Instruct benchmark scores and speedups for conservative and relaxed thresholds.Source: Table 2 · p. 7, §5 · p. 7
  • LLaDA shows similar trends: conservative schedules speed up the base model near parity, while instruction tuning brings higher speedups; fast-decaying settings risk math and long-form degradation.Prophet provides high speedups but underperforms on long-form generation.Source: §5 · p. 7
  • With QPS (γ=4), Dream Instruct’s best SchED setting reaches 4.30 versus Prophet’s 2.91; Dream Base smooth schedules score 1.01–1.12, with Exp-k=16 at 2.03.The QPS table covers all SchED variants and Prophet; higher values are better.Source: Table 3 · p. 7, §5 · p. 7
  • Across five input types, mean per-token entropy decreases monotonically over denoising steps for both Dream Base and Dream Instruct.The figure reports nats per token averaged over evaluation samples, with shaded bands showing one standard deviation.Source: Figure 1 · p. 8

6 ANALYSIS

The analysis tracks predictive entropy across reverse-diffusion steps and finds that instruction-tuned Dream models show faster confidence stabilization than base models.

6 ANALYSIS

  • The analysis tracks per-token predictive entropy on the generated region at each reverse-diffusion step, using model distributions computed from logits.The distributions are conditioned on the prompt and current noised sequence.Source: §6 · p. 8
  • Dream Instruct entropy drops rapidly early and then decreases nearly uniformly; its task trajectories converge to similarly low final entropy despite differing early patterns.Math-heavy prompts start at higher entropy and decay more slowly in the earliest steps than MMLU, HotpotQA, and WMT14.Source: §6 · p. 8
  • Dream Base has more overlapping benchmark trajectories and higher residual entropy overall, aligning with its smaller schedule-based early-exit speedups.The passage describes this as suggesting less decisive posteriors for QA-style inputs.Source: §6 · p. 8

7 DISCUSSION

SchED’s progress-aware stopping is associated with substantial instruction-tuned speedups, while entropy patterns and task differences help explain where acceleration is attainable. Conservative schedules preserve quality more closely than rapidly decaying schedules, and similar patterns appear across model families.

7 DISCUSSION

  • Up to approximately 17× speedups with full baseline accuracy, and average accelerations up to about 4×, are reported for instruction-tuned variants.Dream Base retains higher, flatter entropy trajectories that delay threshold crossing and constrain attainable speedups.Source: §7 · pp. 8–9
  • Instruction tuning’s entropy decline is rapid, whereas Dream Base’s flatter, higher entropy postpones threshold crossing and constrains attainable speedups.For Dream Instruct, entropy falls sharply at first and then decreases approximately uniformly across steps.Source: §7 · pp. 8–9
  • Conservative cosine and linear schedules stay near baseline while reducing refinement steps; γ=4 QPS indicates consistently high quality–efficiency trade-offs.Rapidly decaying exponential schedules with large k and τlow = 0 yield greater acceleration but noticeable accuracy losses.Source: §7 · p. 9
  • Prophet’s localized confidence and discrete commit rule can trigger premature exits on long-form generation, leaving later output underresolved and reducing ROUGE and F1.SchED instead aggregates over the full generated sequence, uses a progress-dependent threshold, and avoids auxiliary suffix prompts.Source: §7 · p. 9
  • Stopping behavior varies by task: multiple-choice GPQA reaches confidence earlier than free-form GSM8K, while long-form tasks benefit from smoothing and full-span aggregation.Translation keeps CHRF close to baseline while reducing steps.Source: §7 · p. 9
  • LLaDA shows analogous behavior: conservative smooth schedules improve efficiency near parity on base variants and yield larger gains on instruction-tuned models.The schedule should match the tolerated error budget: conservative settings favor fidelity, while faster-decaying schedules suit latency-sensitive or fault-tolerant uses that allow modest degradation.Source: §7 · p. 9

8 CONCLUSION

The conclusion presents SchED as a training-free, architecture-agnostic early-exit method for diffusion language models. It reports reduced denoising steps across Dream and LLaDA while preserving performance, with schedule choices spanning fidelity and efficiency trade-offs.

8 CONCLUSION

  • SchED compares a sequence-level confidence statistic with a smooth, progress-conditioned threshold to decide when diffusion decoding should stop.It is training-free and architecture-agnostic, and integrates with existing transfer schedules without modifying the underlying models.Source: §8 · p. 9
  • Across Dream and LLaDA, SchED consistently reduces denoising steps while preserving performance, with particularly large gains on instruction-tuned variants.The conclusion relates those gains to faster uncertainty decay on instruction-tuned variants, as reflected in the entropy analysis and quality–penalized speed metric.Source: §8 · p. 9
  • The schedule family spans quality–efficiency trade-offs, and the paper identifies learned parameters, task-aware aggregation, domain-aware thresholds, and complementary acceleration as future work.The conclusion frames stopping-time decoding as a primitive for deploying dLLMs under accuracy and throughput constraints.Source: §8 · p. 9

LIMITATIONS

The study’s limitations include a task-dependent quality–efficiency trade-off and a narrow inference-time design that leaves several schedule-learning and integration directions unexplored.

LIMITATIONS

  • The appropriate schedule is task-dependent, although experiments select a small set of global configurations for each model and regime.More conservative schedules preserve near-parity quality with moderate speedups; more extreme, rapidly decaying schedules accelerate more but incur marginal average degradation.Source: LIMITATIONS · p. 10
  • SchED treats the schedule as an external inference-time control; joint training, learned aggregation beyond answer-span averaging, and integration with speculative or cache-based denoising remain unexamined.The paper identifies these as directions that could further improve robustness and efficiency.Source: LIMITATIONS · p. 10

A GENERATION CONFIGURATIONS

The configuration table summarizes benchmark-specific decoding settings, including step budgets, answer lengths, shots, and LLaDA block sizes. The settings vary with task type and generation length.

A GENERATION CONFIGURATIONS

  • LLaDA uses a block size of 5 tokens for short MCQ benchmarks and 32 tokens for tasks requiring generation lengths of at least 32 tokens.The block-size setting applies only to LLaDA’s block-diffusion decoder.Source: A GENERATION CONFIGURATIONS · p. 13
  • MCQ tasks use T=5 and very short generation lengths, while math, translation, and long-form tasks use larger budgets of 256–512 and longer outputs.All listed MCQ benchmarks are evaluated by generating the answer region, not by likelihood or ranking of answer options.Source: A GENERATION CONFIGURATIONS · p. 13
  • Table 4 lists decoding hyperparameters per benchmark, with columns covering step budget, maximum answer-region length, in-context shots, and—where applicable—LLaDA block size.Dream uses single-block/suffix refinement and does not require a block-size setting.Source: Table 4 · p. 13, A GENERATION CONFIGURATIONS · p. 13

B ADDITIONAL SchED VARIANTS AND LLaDA RESULTS

The appendix reports Dream Base and Instruct variants alongside full LLaDA Base and Instruct results. Intermediate-curvature exponentials generally sit between slower gradual-decay schedules and faster, higher-curvature schedules in speed and quality.

Appendix scope

  • The appendix adds Dream intermediate-curvature exponentials k∈{4, 8} at thresholds (7.5, 2.5) and (7.5, 0), alongside full LLaDA Base and Instruct results.Other Dream schedule families were already covered in the main text.Source: B ADDITIONAL SchED VARIANTS AND LLaDA… · p. 13

Dream

  • Dream Base Exp–k=8 (7.5, 2.5) reaches 73.53 on MMLU and ties the top WMT14 En–Fr score at 53.26 with ×1.10 mean speedup.Exp–k=4 is also competitive, with 88.63 on PIQA at ×1.07 speedup; Exp–k=16 (7.5, 0) reaches ×2.34 with an expected quality trade-off.Source: B ADDITIONAL SchED VARIANTS AND LLaDA… · pp. 13–14, Table 5 · p. 14
  • On Dream Instruct, Exp–k=4 with (7.5, 2.5) attains the highest overall average score, 58.22, at an average speedup of ×3.97 among the reported schedules (Table 6).Exp–k=8 with (7.5, 2.5) reaches mean speedup ×4.23, matches the MMLU peak (73.53), and yields the top GPQA score (30.80).Source: B ADDITIONAL SchED VARIANTS AND LLaDA… · p. 13

LLaDA

Related papers

Loading 2512.02892v1…