mlsanityData to 7 Oct 2026

LogicDiff: Logic-Guided Denoising Improves Zero-Shot Reasoning in Masked Diffusion Language Models

arXiv 2603.26771 ↗First posted 24 Mar 2026cs.CL cs.LGcited by 1

On the map: LLM reasoning → Diffusion reasoning and test-time scaling

also in 1 more
Abstract (arXiv)

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens from a fully masked sequence. Their standard confidence-based unmasking strategy systematically defers high-entropy logical connective tokens, degrading reasoning performance. We introduce LogicDiff, an inference-time method that replaces confidence-based unmasking with logic-role-guided unmasking. A lightweight classification head (4.2M parameters, 0.05% of the base model) predicts the logical role of each masked position (premise, connective, derived step, conclusion, or filler) from the base model's hidden states with 98.4% accuracy, and a dependency-ordered scheduler unmasks tokens in logical order. In zero-shot settings, LogicDiff improves LLaDA-8B-Instruct accuracy from 22.0% to 60.7% on GSM8K (+38.7 percentage points) and from 23.6% to 29.2% on MATH-500 (+5.6 pp), with less than 6% speed overhead. However, with 8-shot chain-of-thought prompting, the baseline reaches approximately 70% and LogicDiff provides no additional improvement. Analysis reveals that few-shot prompting implicitly resolves the same ordering problem that LogicDiff explicitly addresses, and that fixed role-based ordering can cause premature commitment to numerical values before sufficient context is available. Our results characterize the Flexibility Trap as primarily a zero-shot phenomenon and identify context-adaptive ordering as a key direction for future work.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in Diffusion reasoning and test-time scaling, one of 17 directions in LLM reasoning. Builds on 9 of its 17 references that are in the map. First posted on arXiv 24 Mar 2026.

Builds on

9 of 17 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching26 Feb
  2. 2026The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models21 Jan
  3. 2025Reasoning in Diffusion Large Language Models is Concentrated in Dynamic Confusion Zones19 Nov
  4. 2025Advancing Reasoning in Diffusion Language Models with Denoising Process Rewards2 Oct
  5. 2025d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning16 Apr

Other references · 4

  1. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  2. 2025Large Language Diffusion Modelscited by 872
  3. 2026Autoregressive Models Rival Diffusion Models at ANY-ORDER Generationcited by 3
  4. 2026DOS: Dependency-Oriented Sampler for Masked Diffusion Language Modelscited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Modelscited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2025Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models
  2. 2026Where-to-Unmask: Ground-Truth-Guided Unmasking Order Learning for Masked Diffusion Language Models
  3. 2026Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models
  4. 2025Lookahead Unmasking Elicits Accurate Decoding in Diffusion Language Models
  5. 2026Remask, Don't Replace: Token-to-Mask Refinement in Diffusion Language Models
  6. 2026CAI-DLLM: Convergence Aware Inference for Diffusion Language Models
  7. 2025Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies
  8. 2026NAVIRA: Decoupled Stochastic Remasking for Masked Diffusion Language Models