Masked diffusion language models generate text by starting from a sequence of blank "mask" slots and filling them in over many steps. At each step they decide which slots to reveal and what to put there. Today that decision rests on one number per slot: how confident the model's softmax is at that position. This number is known to mislead in three ways. It reveals boring scaffolding tokens first, which collapses diversity. It lets dependent tokens be revealed together, which produces gibberish. And when a revealed token is flagged as suspicious and re-masked, the model usually writes the same token back.
Speech recognition faced the same overconfidence problem decades ago. Its answer was to stop trusting a single hypothesis and to read confidence off a confusion network: a set of aligned "bins" holding competing word hypotheses, where a word's confidence is the probability mass it keeps after being normalized against its rivals. This note proposes the diffusion analogue. We run a small batch of decoding trajectories ("particles") in parallel. Because a masked diffusion model's positions are fixed, the particles are aligned for free, so every position becomes a confusion bin. At each step we pool the particles' predictions per position. Where they agree, the token is an anchor that can be committed once and shared. Where they disagree, the position is a genuine fork and should stay stochastic. A re-masked token is refilled from the pooled distribution, which contains alternatives that other contexts proposed.
Core Claim
The pooled, cross-particle posterior at a position is a better signal than any single trajectory's softmax, for three purposes: deciding what to commit, deciding how much randomness to inject, and deciding which final answer to trust. This holds even when total compute is matched against running the same particles independently.
Concretely, let $p^{(b)}_i(v)$ be particle $b$'s predicted probability of token $v$ at masked position $i$, given its own partially revealed sequence, for $b=1,\dots,B$. The bin posterior is the mixture $\bar p_i(v)=\frac{1}{B}\sum_b p^{(b)}_i(v)$. We split its uncertainty into two parts:
$$ D_i \;=\; H(\bar p_i) \;-\; \frac{1}{B}\sum_{b} H\big(p^{(b)}_i\big), $$
Here $H$ is Shannon entropy. $D_i$ is the mutual information between the particle index and the token. It is large exactly when each particle is individually sure but the particles disagree, meaning the prediction depends on which context was revealed earlier. That is the signature of a fork, and it is also the signature of a confident local error. A single particle cannot detect either, because its own softmax is sharp. The flip-flop finding in Re-evaluating Confidence Remasking in Masked Diffusion Language Models is this effect in miniature: in 75–95% of re-masked positions, a model conditioned on its own context re-predicts the same token.
The novelty is not "sample several times and vote." It is that disagreement is measured per position during decoding, and it steers three things while decoding is still in progress. The first is shared commitment of anchors. The second is a position temperature set per position from $D_i$ instead of by hand. The third is refilling re-masked slots from $\bar p_i$.
Where This Sits
Speech recognition supplies the method. Evaluating semantic-level confidence scores with multiple hypotheses shows that normalizing an arc's posterior against competitors in its confusion bin produces calibrated confidence. Confidence-based Ensembles of End-to-End Speech Recognition Models and the meta-classifiers in On Addressing Practical Challenges for RNN-Transducer show that multi-hypothesis and confusion-network features beat raw token probabilities for predicting correctness. Accurate and Reliable Confidence Estimation Based on Non-Autoregressive End-to-End Speech Recognition System documents the systematic overconfidence of parallel, non-autoregressive decoders, which is the regime diffusion language models live in. We borrow the bin-normalization idea. We do not borrow full lattices.
The scale-limitation literature is a constraint on the design. Sequence Modeling via Segmentations marginalizes exactly over a segmentation lattice, and that marginalization is what fails to scale to high-variance corpora. We therefore avoid building a combinatorial graph. The structure is one bin per position with $B$ entries, and its cost is linear in $B$.
Four diffusion-decoding papers motivate the work and are its main competitors:
- A Tale of Two Temperatures diagnoses the anchor-fork degeneracy and fixes it with a hand-tuned global position temperature $T_{pos}$. It also notes that the added diversity can hurt majority voting. We replace the global $T_{pos}$ with a per-position one driven by $D_i$, and we replace voting with bin-posterior scoring.
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (the EB-Sampler) bounds a heuristic single-trajectory entropy to decide how many tokens to reveal at once. We test whether the mixture entropy $H(\bar p_i)$ is a tighter proxy for joint-dependence risk.
- Lookahead Unmasking Elicits Reliable Decoding in Diffusion Language Models (LookUM) scores a few candidate unmasking sets by the entropy of a one-step simulated future. It degrades beyond depth one. Our particles spend compute on width, meaning parallel contexts, rather than depth.
- Semantic Uncertainty also measures agreement across samples, but only on finished outputs and only to flag hallucinations. We measure agreement mid-generation and act on it.
Research Program
The sampler maintains $B$ particles (we plan $B\in\{2,4,8\}$) on LLaDA-8B-Instruct and Dream-7B, the models used by the competing papers. Each step proceeds in three parts.
- Pool and commit anchors. Run one batched forward pass per particle. Compute $\bar p_i$ and $D_i$ for every masked position. Positions whose mixture mass on the top token exceeds a threshold, and whose $D_i$ is near zero, are committed identically in every particle. This is the shared-anchor move. Its batch size is bounded by an entropy-bound rule applied to $H(\bar p_i)$.
- Diversify forks. Among the remaining positions, each particle samples which to reveal using a position temperature that increases with $D_i$. Genuine forks are therefore resolved differently across particles, and low-context positions are not revealed early.
- Refill from the pool. Optionally, a committed token whose $D_i$ later spikes is re-masked. It is resampled from $\bar p_i$ restricted to tokens other than the one just removed, provided that token keeps more than a minimum share of the mixture mass.
At the end, each particle's answer is scored by the average bin posterior of its committed tokens. This mirrors word-level confusion-network confidence in speech recognition.
The analysis side asks three questions. Is $\bar p_i$ better calibrated than single-particle softmax at predicting token-level errors? Does $D_i$ localize forks and errors? And does pool refilling reduce the flip-flop rate? Sudoku is the key instrument for the first two, because every cell has a checkable ground truth. EB-Sampler already reaches near-perfect Sudoku accuracy at 10–15 function evaluations, so Sudoku is useful for measuring calibration, not for comparing accuracy. LookUM's arithmetic diagnostic, where injected local errors cause entropy spikes, tests whether $D_i$ fires earlier and more selectively than plain entropy.
All compute is reported as total forward passes, meaning steps times particles, and also as wall-clock time on one GPU, since batching makes particles partly free in latency but not in FLOPs.
Baselines
Naive reference. The do-nothing reference is the models' default greedy low-confidence unmasking with one trajectory. It is run at the same block length of 32. At BL=32, Re-evaluating Confidence Remasking shows that remasking gains shrink to about 0.4–0.5%, so gains measured at large block lengths would be illusory. Fast-dLLM, the no-remasking parallel schedule evaluated there, runs alongside it.
Efficiency competitor. The strongest efficiency competitor is the EB-Sampler. Its $\gamma$ threshold is tuned on a held-out split with the same tuning budget we give our thresholds. It reports a 2x–3x reduction in function evaluations without losing pass@1 relative to single-token unmasking. Beating it means matching its pass@1 on GSM8K, HumanEval and MBPP at fewer total forward passes across all particles. Merely matching it at fewer steps per particle does not count.
Accuracy competitor. The strongest accuracy competitor is LookUM with 2–3 candidate paths and depth one, its best reported setting. It reports up to 4-point gains on GSM8K and HumanEval and 8 points on MBPP over confidence unmasking, and it brings base LLaDA to the level of RL-tuned LLaDA 1.5. The file gives only these improvements, not absolute scores, so we will reproduce its baseline and report both. LookUM will be given the same total forward-pass budget as our $B$-particle sampler, including its lookahead passes. Beating it means higher pass@1 at equal budget.
Diversity competitor. For pass@k and diversity, the competitor is TCT from A Tale of Two Temperatures, run $B$ times independently with $T_{pos}$ grid-searched per task. The paper reports 70% on MATH at about 4,000 function evaluations, versus about 8,000 for autoregressive sampling. Beating it means equal or better pass@k at equal budget with no per-task $T_{pos}$ search. It also means our bin-posterior answer selection must beat TCT plus majority voting at pass@1.
Excluded as out of regime. A2D2 reports GSM8K Pass@1 rising from 35.71% to 60.96% at 128 steps. It requires reward fine-tuning of an any-length model. We cite it as a training-based ceiling, not as a matched training-free competitor.
Ablations. Three ablations isolate the key ingredient:
- Independent particles with bin-posterior scoring only at the end. This is self-consistency with a smarter vote and no mid-decoding sharing.
- Shared anchors with a fixed global $T_{pos}$.
- The EB bound computed on $H(\bar p_i)$ versus on a single particle's entropy.
Why not just…?
Why not just run independent samples and take a majority vote? This is the most serious threat, and the first ablation exists to settle it. If sharing and per-position temperature add nothing over end-of-decoding aggregation, the idea reduces to self-consistency, and the proposal should shrink to a better answer-scoring rule. There is evidence the threat is incomplete: A Tale of Two Temperatures reports that diverse samples can dilute the majority mode. Mid-decoding commitment of consensus anchors should also cut joint-dependence errors, which voting cannot repair after the fact. We will report the pass@$B$ oracle so readers can see how much headroom any selection rule has.
Why not just widen LookUM to $B$ candidates? LookUM compares candidates by the entropy of a simulated future from a single context. It never observes how predictions at a position change across different revealed contexts, which is what $D_i$ measures. Its known failure is that simulated futures become noisy beyond depth one. The matched-budget comparison in which LookUM gets $k=B$ candidates answers this directly. If LookUM with $k=B$ matches us, our contribution narrows to calibration and remasking.
Why not just calibrate the single-trajectory softmax, for example with temperature scaling or a TempNet-style learned temperature? To Cool or not to Cool? (TempNet) learns instance-level temperatures for autoregressive models, and plain temperature scaling rescales confidence monotonically. Neither can reorder positions whose errors are invisible from the model's own context. The flip-flop rate suggests these errors are exactly the ones that matter. The Sudoku calibration experiment includes a temperature-scaled softmax baseline fit on held-out puzzles. If it matches $\bar p_i$ on AUROC for predicting wrong cells, the multi-particle machinery is unjustified.
Why not just remask more aggressively, as WINO does? The re-evaluation paper already shows WINO gives about 0.4–0.5% at BL=32 while collapsing diversity, and that flip-flops dominate. Our refilling rule changes the source of the replacement token rather than the trigger for remasking. The test is simple: the flip-flop rate under pool refilling versus under WINO, on identical flagged positions.
Decisive Experiments
The first check is calibration. On Sudoku, and on GSM8K using per-step tokens later contradicted by the final committed answer, we compare three predictors of wrong committed tokens: single-particle softmax, temperature-scaled softmax, and the bin posterior. We report expected calibration error and AUROC. If the bin posterior does not clearly win at $B=4$, the project stops there.
The second check is the matched-budget accuracy and efficiency curve. It plots pass@1 on GSM8K, MATH-500, HumanEval and MBPP against total forward passes for all baselines above. Code is evaluated with the mxeval harness (amazon-science/mxeval). The claim holds only if our curve lies above both EB-Sampler and budget-matched LookUM somewhere in the practical range.
The third check covers diversity and the flip-flop rate. It measures pass@k against independent TCT without per-task tuning, and it measures the flip-flop rate under pool refilling. We expect the latter to fall far below the 75–95% reported for WINO. If flip-flops persist even when alternatives come from other particles, that supports the re-evaluation paper's view that training-free decoding cannot supply missing knowledge. We would report that as a negative result, bounded by the pass@$B$ oracle.
References
- Re-evaluating Confidence Remasking in Masked Diffusion Language Models
- Evaluating semantic-level confidence scores with multiple hypotheses
- Confidence-based Ensembles of End-to-End Speech Recognition Models
- On Addressing Practical Challenges for RNN-Transducer
- Accurate and Reliable Confidence Estimation Based on Non-Autoregressive End-to-End Speech Recognition System
- Sequence Modeling via Segmentations
- A Tale of Two Temperatures: Simple, Efficient, and Diverse Sampling from Diffusion Language Models
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking
- Lookahead Unmasking Elicits Accurate Decoding in Diffusion Language Models
- Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
- A2D2: Finetuning Any-Length Discrete Diffusion for Adaptive Decoding
- To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO