Where and When to Commit: Candidate-Aware Decoding for Diffusion Language Models

• URL: http://arxiv.org/abs/2607.28166v1
• Date: Thu, 30 Jul 2026 13:04:47 GMT
• Title: Where and When to Commit: Candidate-Aware Decoding for Diffusion Language Models
• Abstract summary: Diffusion language models expose a provisional prediction at every denoising step.<n>Generation-time early exit stops decoding before the schedule is exhausted.<n>We introduce a training-free, candidate-aware early-exit framework that keeps the two axes separate and matches each decision to evidence of its own scope.
• Score: 29.849017661114342
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Diffusion language models (DLMs) expose a provisional prediction at every denoising step, creating an opportunity for generation-time early exit that stops decoding before the schedule is exhausted. Existing early-exit gates decide termination from fixed-region confidence statistics or schedule-dependent rules, evidence too coarse for a decision that freezes every remaining position at once, so they fire prematurely on long chain-of-thought outputs whose answers stabilize only near the end. Adaptive sampling, the other axis of training-free acceleration, paces how quickly positions commit while decoding continues but never verifies that the output itself has stabilized. We introduce a training-free, candidate-aware early-exit framework that keeps the two axes separate and matches each decision to evidence of its own scope. Confidence-Verified Commit (CVC) governs when the sequence may stop by verifying confidence and sustained argmax stability over the dynamically extracted candidate span using a deterministic parser specified from each task's output format. Block-Wise Early Commit (BWEC) governs where to accelerate by applying a cheaper local rule to non-final blocks, while leaving the final block and global termination under CVC. We refer to their combination as LATCH (Localized Acceleration with Tracked-Candidate Halting). Unlike prior methods, LATCH needs no suffix-prompt construction; it is prompt-anchor-free but format-aware. We evaluate LATCH end to end on 11 tasks under zero-shot settings using LLaDA and Dream. LATCH stays within 2.0 percentage points of full-decoding accuracy across all 22 evaluation settings, with one frozen hyperparameter set that transfers cross-backbone untuned, while achieving end-to-end TPS speedups of 9.3-17.8x on short-answer tasks and 2.0-3.3x on long-reasoning tasks.

Related papers

• Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models [66.32132912853372]
  Diffusion Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive language models.<n>We propose Ripple-Pivot Search (RPS), a training-free decoding method that seeks mid-entropy positions as promising candidate pivots.<n>RPS achieves 4-10$times$ wall-clock speedup over the standard decoder while preserving generation quality, and improves accuracy over the previous lookahead baseline by up to 5.49%.
  arXiv  Detail & Related papers  (2026-08-12T07:32:29Z)
• CORA-Diff: Confidence-Oriented Residual Acceptance for Efficient Diffusion Language Model Inference [24.176666448303095]
  Diffusion language models update many tokens in parallel, yet practical decoders often use a fixed denoising horizon.<n>We propose CORA-Diff, a training-free method that preserves the original transfer rule and applies confidence-and-persistence gating only to positions that rule leaves unresolved.<n>Results show that native confidence and persistence enable reliable residual acceptance, reducing repeated denoising while preserving task quality.
  arXiv  Detail & Related papers  (2026-07-31T07:32:10Z)
• TACG: Trajectory-Aware Commit Gating for Diffusion Language Model Decoding [38.72024551312125]
  Trajectory-Aware Commit Gating (TACG) is a gate-level decoder that anchors token identities to the base posterior.<n>We evaluate TACG on LLaDA, Dream, and LLaDA2-Mini benchmarks.
  arXiv  Detail & Related papers  (2026-07-03T11:45:21Z)
• Efficient Diffusion LLMs via Temporal-Spatial Parallel Decoding and Confidence Extrapolation [15.129362213838974]
  We propose a trace-aware decoding framework with two components.<n>First, Temporal-Spatial Parallel Decoding (TSPD) uses a lightweight temporalspatial controller that consumes per-token trajectory features, including confidence, entropy, and momentum, together with token position, to decide when a token has converged and can be safely fixed.<n>Second, we introduce Confidence Extrapolation (CE), a training-free state-space module that forecasts future logit trends with uncertainty to support proactive decisions.
  arXiv  Detail & Related papers  (2026-05-29T02:29:28Z)
• The Path Matters: Learning a Token-Commitment Policy for Diffusion Language Models [52.93186090124315]
  We argue that token commitment can be learned as a reusable trace-state policy.<n>We introduce TraceLock, a lightweight plug-in controller that instantiates this policy for a frozen diffusion language model.
  arXiv  Detail & Related papers  (2026-05-23T18:23:46Z)
• STDec: Spatio-Temporal Stability Guided Decoding for dLLMs [49.55447757907809]
  Large Diffusion Language Models (dLLMs) have achieved rapid progress, viewed as a promising alternative to the autoregressive paradigm.<n>Most dLLM decoders still adopt a global confidence threshold, and do not model local context from neighboring decoded states or temporal consistency of predicted token IDs.<n>We observe strong stability neartemporal stability in dLLM decoding, and our approach includes spatial-aware decoding and temporal-aware decoding.
  arXiv  Detail & Related papers  (2026-04-07T18:13:31Z)
• DenoiseFlow: Uncertainty-Aware Denoising for Reliable LLM Agentic Workflows [20.319113495948294]
  We formalize the multi-step reasoning process as a Noisy MDP.<n>We propose DenoiseFlow, a closed-loop framework that performs progressive denoising through three coordinated stages.
  arXiv  Detail & Related papers  (2026-02-28T08:11:38Z)
• Don't Let It Fade: Preserving Edits in Diffusion Language Models via Token Timestep Allocation [9.670831676453673]
  We describe update forgetting, in which uniform and context updates induce token level fluctuations across timesteps.<n>We propose Token Timestep Allocation (TTA), which realizes soft and semantic token ordering via per token timestep schedules.
  arXiv  Detail & Related papers  (2025-10-30T07:21:05Z)
• Pipeline Parallelism is All You Need for Optimized Early-Exit Based Self-Speculative Decoding [73.67253077506672]
  Large language models (LLMs) deliver impressive generation quality, but incur very high inference cost.<n>Early-exit based self-speculative decoding (EESD) has emerged to mitigate this cost.<n>We propose Pipeline-Parallel Self-Speculative Decoding (PPSD) that fully pipelines the draft and verification work.
  arXiv  Detail & Related papers  (2025-09-19T04:51:41Z)
• Diffusion Language Models Know the Answer Before Decoding [56.96815863705218]
  Diffusion language models (DLMs) have emerged as an alternative to autoregressive approaches.<n>Our work highlights and leverage an overlooked property of DLMs early answer convergence.<n>We introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding.
  arXiv  Detail & Related papers  (2025-08-27T15:40:25Z)
• Acting in Delayed Environments with Non-Stationary Markov Policies [57.52103323209643]
  We introduce a framework for learning and planning in MDPs where the decision-maker commits actions that are executed with a delay of $m$ steps. We prove that with execution delay, deterministic Markov policies in the original state-space are sufficient for attaining maximal reward, but need to be non-stationary. We devise a non-stationary Q-learning style model-based algorithm that solves delayed execution tasks without resorting to state-augmentation.
  arXiv  Detail & Related papers  (2021-01-28T13:35:37Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.