The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models

• URL: http://arxiv.org/abs/2605.29123v1
• Date: Wed, 27 May 2026 21:33:37 GMT
• Title: The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models
• Abstract summary: We argue that confidence-based decoding is inherently misaligned with the logical-flow trajectories required for complex reasoning.<n>We show that confidence-aligned training actively entrenches this misalignment.<n>In contrast, random masking -- despite its perceived inefficiency -- robustly preserves the reasoning-trajectory conditionals essential for solving the challenging tail.
• Score: 5.43345665278304
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Masked diffusion language models (MDMs) uniquely support any-order generation, with confidence-based decoding currently serving as the de facto standard inference policy. To optimize for this, recent training schemes attempt to align training mask patterns directly with those observed during generation. However, we argue that confidence-based decoding is inherently misaligned with the logical-flow trajectories required for complex reasoning, and that confidence-aligned training actively entrenches this misalignment. We make this concrete using multi-digit addition, where the decoding strategy prematurely predicts locally easy digits before resolving their long-range dependencies, producing high-confidence errors on challenging inputs. While traditional random masking keeps the failure rate low on this challenging tail, confidence-aligned training amplifies the error rate by an order of magnitude. Across five distinct reasoning tasks, this same pattern emerges with task-dependent severity: confidence-based decoding induces failures on highly complex inputs, and confidence-aligned training exacerbates them. In contrast, random masking -- despite its perceived inefficiency -- robustly preserves the reasoning-trajectory conditionals essential for solving the challenging tail.

Related papers

• Escaping Confidence Trap: Evolutionary Decoding for Mathematical Reasoning in Diffusion LLMs [69.52853771994451]
  We analyze the decoding trajectories of LLaDA 2.0 and identify a recurring diffusion confidence trap.<n>We propose Evolutionary Decoding, a training-free test-time scaling framework that views diffusion decoding as an evolutionary process.
  arXiv  Detail & Related papers  (2026-08-01T11:48:25Z)
• Focus When Necessary: Adaptive Routing and Collaborative Grounding for Training-Free Visual Grounding [56.38477453373368]
  LazyMCoT is a dynamic and training-free framework that adaptively allocates visual grounding efforts based on sample difficulty.<n>LazyMCoT rivals training-based approaches by simultaneously improving reasoning accuracy and reducing average inference latency.
  arXiv  Detail & Related papers  (2026-06-15T03:17:44Z)
• Re-evaluating Confidence Remasking in Masked Diffusion Language Models [13.69108579131071]
  We revisit the empirical evaluation of a representative post-hoc remasking method, WINO.<n>We find that while confidence-based remasking can mitigate errors, it also exacerbates the diversity collapse previously reported for confidence-based unmasking.
  arXiv  Detail & Related papers  (2026-06-10T15:41:26Z)
• Locally Confident, Globally Stuck: The Quality-Exploration Dilemma in Diffusion Language Models [52.61023005303122]
  We show that low-confidence remasking improves a proxy for quality while constraining the entropy of the induced sequence distribution.<n>We develop a simple Independent--Hastings sampler that approximately targets this distribution during decoding.
  arXiv  Detail & Related papers  (2026-04-01T02:01:30Z)
• Thinking by Subtraction: Confidence-Driven Contrastive Decoding for LLM Reasoning [58.331709210563616]
  Thinking by Subtraction is a confidence-driven contrastive decoding approach.<n>A small subset of low-confidence tokens disproportionately contributes to reasoning errors and unnecessary output expansion.<n>Our method, Confidence-Driven Contrastive Decoding, detects low-confidence tokens during decoding and intervenes at these positions.
  arXiv  Detail & Related papers  (2026-02-20T14:13:22Z)
• CORE: Context-Robust Remasking for Diffusion Language Models [51.59514489363897]
  We propose Context-Robust Remasking (CORE), a training-free framework for inference-time revision.<n>Rather than trusting static token probabilities, CORE identifies context-brittle tokens by probing their sensitivity to targeted masked-context perturbations.<n>On LLaDA-8B-Base, CORE delivers consistent improvements across reasoning and code benchmarks, outperforming compute-matched baselines and improving MBPP by up to 9.2 percentage points.
  arXiv  Detail & Related papers  (2026-02-04T00:12:30Z)
• Reflective Confidence: Correcting Reasoning Flaws via Online Self-Correction [14.164508061248775]
  Large language models (LLMs) have achieved strong performance on complex reasoning tasks using techniques such as chain-of-thought and self-consistency.<n>We propose reflective confidence, a novel reasoning framework that transforms low-confidence signals from termination indicators into reflection triggers.<n> Experiments on mathematical reasoning benchmarks, including AIME 2025, demonstrate significant accuracy improvements over advanced early-stopping baselines at comparable computational cost.
  arXiv  Detail & Related papers  (2025-12-21T05:35:07Z)
• Network Inversion for Generating Confidently Classified Counterfeits [11.599035626374409]
  In vision classification, generating inputs that elicit confident predictions is key to understanding model behavior and reliability.<n>We extend network inversion techniques to generate Confidently Classified Counterfeits (CCCs)<n>CCCs offer a model-centric perspective on confidence, revealing that models can assign high confidence to entirely synthetic, out-of-distribution inputs.
  arXiv  Detail & Related papers  (2025-03-26T03:26:49Z)
• Selective Learning: Towards Robust Calibration with Dynamic Regularization [79.92633587914659]
  Miscalibration in deep learning refers to there is a discrepancy between the predicted confidence and performance. We introduce Dynamic Regularization (DReg) which aims to learn what should be learned during training thereby circumventing the confidence adjusting trade-off.
  arXiv  Detail & Related papers  (2024-02-13T11:25:20Z)
• Addressing Mistake Severity in Neural Networks with Semantic Knowledge [0.0]
  Most robust training techniques aim to improve model accuracy on perturbed inputs. As an alternate form of robustness, we aim to reduce the severity of mistakes made by neural networks in challenging conditions. We leverage current adversarial training methods to generate targeted adversarial attacks during the training process. Results demonstrate that our approach performs better with respect to mistake severity compared to standard and adversarially trained models.
  arXiv  Detail & Related papers  (2022-11-21T22:01:36Z)

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.