The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models
On the map: LLM reasoning → Diffusion reasoning and test-time scaling
also in 1 more
Abstract (arXiv)
Chain-of-thought reasoning helps autoregressive models solve complex problems by generating intermediate steps that support later predictions. Masked diffusion models (MDMs) offer a similar opportunity through arbitrary-order generation: they can ideally reveal intermediate results along logical dependencies. In practice, however, standard decoding simply prioritizes high-confidence tokens, which need not align with this dependency order. We identify this discrepancy as the confidence shortcut: models commit with high certainty to plausible tokens while neglecting long-range dependencies. In multi-digit addition, models predict higher-order digits without properly tracking carries through long chains. Controlled pretraining across diverse reasoning tasks confirms that confidence-guided ordering often selects suboptimal sequences, and confidence-aligned training schemes can exacerbate these failures---for example, increasing addition error rates by an order of magnitude. Our findings caution against relying solely on confidence to choose generation orders and against training objectives that reinforce this preference. The experimental code is available at https://github.com/jinha2536/mdm-arithmetic.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Diffusion reasoning and test-time scaling, one of 17 directions in LLM reasoning. Builds on 41 of its 47 references that are in the map. First posted on arXiv 27 May 2026.
Builds on
41 of 47 references.
Closest ancestors · 7
Same direction, newest first
- 2026LogicDiff: Logic-Guided Denoising Improves Zero-Shot Reasoning in Masked Diffusion Language Models24 Mar
- 2026Think First, Diffuse Fast: Improving Diffusion Language Model Reasoning via Autoregressive Plan Conditioning20 Feb
- 2025d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation25 Sep
- 2025Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models15 May
- 2025d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning16 Apr
- 2025Remasking Discrete Diffusion Models with Inference-Time Scaling1 Mar
- 2024Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning18 Oct
Other references · 34
- 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
- 2021Evaluating Large Language Models Trained on Codecited by 8,320
- 2020Shortcut Learning in Deep Neural Networkscited by 1,969
- 2021Structured Denoising Diffusion Models in Discrete State-Spacescited by 1,719
- 2022MaskGIT: Masked Generative Image Transformercited by 1,099
- 2025Large Language Diffusion Modelscited by 912
- 2024Simple and Effective Masked Diffusion Language Modelscited by 848
- 2023Discrete Diffusion Modeling by Estimating the Ratios of the Data Distributioncited by 745
- 2024Simplified and Generalized Masked Diffusion for Discrete Datacited by 626
- 2025Dream 7B: Diffusion Large Language Modelscited by 544
- 2022A Continuous Time Framework for Discrete Denoising Modelscited by 451
- 2024Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Datacited by 381
22 more
- 2025Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusionscited by 214
- 2025DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generationcited by 187
- 2025Hierarchical Reasoning Modelcited by 177
- 2025LLaDA2.0: Scaling Up Diffusion Language Models to 100Bcited by 166
- 2024Stream of Search (SoS): Learning to Search in Languagecited by 145
- 2018ListOps: A Diagnostic Dataset for Latent Tree Learningcited by 129
- 2023Teaching Arithmetic to Small Transformerscited by 123
- 2024Transformers Can Do Arithmetic with the Right Embeddingscited by 78
- 2025Fine-Tuning Masked Diffusion for Provable Self-Correctioncited by 39
- 2024Causal Language Modeling Can Elicit Search and Reasoning Capabilities on Logic Puzzlescited by 38
- 2023A Configurable Library for Generating and Manipulating Maze Datasetscited by 31
- 2025Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policiescited by 22
- 2026Confidence-Based Decoding is Provably Efficient for Diffusion Language Modelscited by 18
- 2025Planner Aware Path Learning in Diffusion Language Models Trainingcited by 14
- 2025Corrective Diffusion Language Modelscited by 11
- 2025Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMscited by 10
- 2026Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Trainingcited by 10
- 2025Seemingly Simple Planning Problems are Computationally Challenging: The Countdown Gamecited by 9
- 2026Where-to-Unmask: Ground-Truth-Guided Unmasking Order Learning for Masked Diffusion Language Modelscited by 9
- 2026Unifying Masked Diffusion Models with Various Generation Orders and Beyondcited by 6
- 2025Reinforced Context Order Recovery for Adaptive Reasoning and Planningcited by 4
- 2024Show Me How It's Done: The Role of Explanations in Fine-Tuning Language Modelscited by 3
Built on by
Citing papers per year: 2026: 1.
1 paper.
Similar, not cited
Close in wording, no citation either way
- 2026CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models
- 2026Escaping Confidence Trap: Evolutionary Decoding for Mathematical Reasoning in Diffusion LLMs
- 2025Lookahead Unmasking Elicits Accurate Decoding in Diffusion Language Models
- 2026Who Should Lead Decoding Now? Tracking Reliable Trajectories for Ensembling Masked Diffusion Language Models
- 2026When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models
- 2026CARVE: Verified Expansion for Variable-Length Generation in Diffusion Language Models
- 2026CORE: Context-Robust Remasking for Diffusion Language Models
- 2026Re-evaluating Confidence Remasking in Masked Diffusion Language Models