Splitting masked-diffusion decoding into five axes, researchers find state adaptation pays off only in a few predictable states, and selective intervention lifts task-macro utility by 3.2, 4.0, and 5.6 points across three models

The paper organizes MDM inference into a five-axis decision space (score, cardinality, region, commitment, planning) and uses Theorem 1's reversal decomposition to write adaptation value as reversal frequency times average available utility margin, showing that a small global adaptation gap can arise from rare reversals, small margins, or both. Prior work proposed individual rules such as confidence, entropy, KL stability, block or dilated regions, remasking, and look-ahead search, but rarely studied when the preferred choice should change; this work maps existing methods onto a common set of axes and gives a structural characterization of when adaptation value is nonzero. Theorem 1 provides a formal decomposition and lower bound; the appendix maps LLaDA, Fast-dLLM, KLASS, DUS, ReMDM, WINO, Info-Gain and others onto the five-axis table, making this a conceptual and theoretical contribution.

Across three models and ten tasks, adaptation opportunity is highly heterogeneous: cross-fitted candidate-set oracle gaps range roughly 0 to 0.0454 for LLaDA-8B, 0 to 0.0450 for LLaDA-1.5, and 0 to 0.1325 for Dream, the strongest axis changes across models, and fixed actions remain competitive in many settings. Rather than assuming adaptive decoding is better, the work quantifies opportunity per task and per axis with cross-fitted candidate-set oracle gaps, showing gains concentrate in structured tasks such as Constrained JSON Fill, Unique List Commit, HTML Close Tags, and Carry RTL. Up to 100 evaluation prompts per task (68 for JSON Mode Eval), eight spaced states per prompt, four continuation rollouts per state-action pair, and two-fold rollout cross-fitting for candidate-set oracle quantities; prompt-cluster bootstrap 95% intervals are reported in the appendix.

Existence of oracle opportunity does not imply exploitability: model-level mean AUROC is approximately 0.5 for LLaDA-8B, 0.5 for LLaDA-1.5, and 0.6 for Dream, and only a few task-axis settings show positive selective lift; for instance Dream JSON Mode Eval/score reaches AUROC 0.891 yet captures 0% of oracle utility mass in the top 10%, showing binary detection and utility-relevant ranking are distinct. The paper explicitly separates action selection (an axis diagnostic proposes a candidate action) from opportunity detection (a validation-calibrated detector decides whether to deviate from the fixed action), instead of assuming one adaptive controller solves both. The detector is calibrated only on validation prompts via quantile bins and never sees held-out rollout utility; the paper reports AUROC, Spearman correlation, positive-state precision/recall, opportunity capture across coverage, and random-coverage and oracle-gate controls.

Selective adaptation yields gains at both the one-step and end-to-end levels: on LLaDA-8B Constrained JSON Fill/region, the top 5%, 10%, and 20% of detector-ranked states capture 35.3%, 56.9%, and 84.3% of oracle opportunity, with realized one-step lift saturating at +0.0594 by 10% coverage; end-to-end selective decoding raises task-macro utility from 0.392 to 0.424 (LLaDA-8B), 0.344 to 0.384 (LLaDA-1.5), and 0.411 to 0.467 (Dream), while always-diagnostic adaptation is weaker overall (0.370, 0.334, 0.447). This supports a conditional statement: when opportunity is predictable and the action diagnostic is useful, selective one-step adaptation can recover substantial local utility at limited coverage, and the benefit comes from deciding when to intervene rather than adapting unconditionally. Coverage-dependent tables report five coverage levels (5%/10%/20%/50%/100%) with random-coverage and oracle-gate controls; end-to-end results are reported on a pre-specified ten-task suite, with selective adaptation improving 3/10 tasks on LLaDA-8B, 4/10 on LLaDA-1.5, and 3/10 on Dream, and degrading 0, 1, and 1 tasks respectively.

The results concern inference-time decoding for masked diffusion language models, in settings where states can be sampled, one-step utility estimated, and validation prompts are available; the paper holds the planning axis fixed and studies only the four transition-level axes (score, cardinality, region, commitment), changing one axis at a time. It primarily serves researchers and practitioners who want to improve decoding without retraining, and the intended setting is structured or constraint-heavy generation such as Constrained JSON Fill, Unique List Commit, HTML Close Tags, and Carry RTL, where opportunity is concentrated and predictable.

Detectors are near chance in most settings, with model-level mean AUROC around 0.5, 0.5, and 0.6, and only a few task-axis settings show positive selective lift, so which regimes are predictable remains an open question. Some end-to-end tasks show no gain or degradation (for example LLaDA-1.5 Multi-span Cloze at -0.040 and Dream Unique List Commit at -0.028), so the link between one-step utility and terminal quality needs more task-level validation. The paper also holds planning fixed and changes one axis at a time, leaving open whether joint multi-axis adaptation would do better; this summary is based on the full text without auditing every appendix table and figure, so exact numbers should be checked against the original appendix.