Unlike autoregressive models, which generate tokens sequentially and benefit from reasoning-before-answering strategies such as Chain-of-Thought, Masked Diffusion Language Models (MDLMs) refine all sequence positions simultaneously, raising questions about how these models handle tasks requiring justified verdicts. In this work, we investigate the dynamics of MDLM reasoning on fact verification, examining whether justifications serve as genuine reasoning or post-hoc rationalization. We observe that MDLMs typically converge on a verdict early in the diffusion process, treating it as a global anchor that is resolved before the justification is complete. Crucially, enforcing a reasoning-first constraint via delayed verdict unmasking actively degrades performance, dropping accuracy from 86.2% to 71.9% as accumulating justification tokens introduce inconsistencies that override initially correct predictions. Interventional experiments reveal that the model rationalizes incorrect forced verdicts in 56% of cases, and that verdicts are strongly causally dependent on justification quality (57.3% accuracy with corrupted justifications vs. 97.1% with ground-truth). This causal dependence explains the degradation under forced deliberation: as the model generates noisy justification tokens, it conditions on them, gradually overriding its initially correct assessment. Our findings suggest that for fact verification with MDLMs, extended deliberation can be counterproductive, risking the dilution of accurate early predictions with noise introduced during justification generation.
Figures & tables
Figure 1: An example of refinement drift. When forced to generate most of the justification (regardless of verdict confidence) before predicting the verdict, the model reaches the wrong conclusion.
Model
Output Order
Acc (%)
LLaMA 3.1 (8B)
VJ
69.7
LLaDA-8B
VJ
86.2
Qwen3-8B
VJ
89.5
LLaMA 3.1 (8B)
JV
72.5
Qwen3-8B
JV
88.0
LLaDA-8B
JV
89.0
Table 1: Performance of baseline models on AVeriTeC development set. Order indicates the order the justification ( J ) and verdict ( V ) are generated in.
Model
Deliberation (%)
Acc (%)
LLaDA-8B
0
86.2
LLaDA-8B
25
86.6
LLaDA-8B
50
79.2
LLaDA-8B
75
73.8
LLaDA-8B
90
71.9
Table 2: Performance impact of forced deliberation on LLaDA. Deliberation % is calculated based on the original span of 64 reasoning tokens.
Claim
Evidence
Truth
Pred.
Generated Explanation & Error
Amy Coney Barrett was confirmed as Supreme Court Justice on October 26, 2020.
Q: Is Amy Coney Barrett confirmed as Supreme Court Justice? A: Yes.
Supp.
Ref.
“…the date of confirmation is incorrect. She was confirmed on September 26, 2020.” Hallucinates wrong date; evidence has no dates.
Jammu and Kashmir was removed from the UN list of ‘unresolved disputes’.
Q: Was J&K an item on the Security Council agenda? A: “J&K has never been an item on the agenda… it is listed in a separate category.”
Ref.
Supp.
“The claim is true because the dispute has never been an item on the agenda… indicating it has been removed.” Conflates “never listed” with “removed.”
Table 3: Example predictions from the Reliance Test showing how corrupted justifications mislead the model. Supp. = Supported, Ref. = Refuted.
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.
Shaik Aman
Independent Researcher, Nellore, Andhra Pradesh, India
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.
Dueun Kim, Albert No
Department of Artificial Intelligence, Yonsei University
Masked diffusion language models (MDLMs) can generate text efficiently by predicting multiple masked tokens in parallel, but predictions from the same forward pass are not necessarily reliable when committed together. We study when parallel commitment is reliable. Our diagnostics show that confidence alone does not determine a reliable commitment order: confident predictions near the end of the sequence can fix an answer before its supporting computations are established, and downstream predictions become less reliable as the uncertainty of their upstream context grows. At the same time, a single forward pass can already resolve several masked tokens, and predictions that remain stable across the final layers are more likely to be correct. Based on these findings, we propose Reliable Parallel Decoding (RPD), a training-free method that selects candidates by layerwise prediction stability and final confidence, and commits them under a cumulative entropy budget over their preceding masked positions. RPD defers predictions with uncertain upstream context while committing the remaining candidates in parallel, without relying on a fixed block schedule. Across mathematical reasoning and code generation benchmarks on LLaDA and Dream, RPD achieves the highest decoding throughput among the evaluated methods while maintaining or improving accuracy.
Zhenghao He, Bohan Liu, Guangzhi Xiong +1
Department of Computer Science, University of Virginia