PUNT Sampler Introduced to Balance Independence and Confidence The new PUNT sampler identifies token dependencies within masked diffusion models and removes lower‑confidence tokens from conflicting groups, ensuring that selected unmasking indices satisfy approximate conditional independence while prioritising high‑confidence predictions [1]. This design directly addresses the trade‑off that has limited parallel sampling in prior approaches [1]. By structuring token groups this way, PUNT maintains coherence across simultaneously generated tokens [1].

Parallel Unmasking Achieves Faster Inference Without Accuracy Loss Enforcing conditional independence lets PUNT update many tokens at once, delivering inference speeds markedly higher than traditional left‑to‑right autoregressive generation [1]. Experiments show that this parallel unmasking does not sacrifice generation quality, matching or exceeding sequential baselines on standard metrics [1]. The speed advantage becomes more pronounced for longer sequences, where sequential models suffer latency bottlenecks [1].

Benchmark Results Show Up to 16% Accuracy Gain on IFEval On the IFEval benchmark, PUNT outperforms strong training‑free baselines, delivering up to a 16 % increase in accuracy [1]. The improvement holds even when compared to one‑by‑one sequential generation for extended texts [1]. These results indicate that parallel generation can be both faster and more accurate when guided by PUNT’s confidence‑driven selection [1].

Robustness Reduces Hyperparameter Tuning and Reveals Hierarchical Planning Performance gains persist across a wide range of hyperparameter settings, suggesting that PUNT lessens reliance on brittle tuning required by earlier methods [1]. Observations reveal an emergent hierarchical generation pattern: the sampler first establishes high‑level paragraph structure before refining local details, resembling a planning process [1]. This behavior contributes to the model’s strong alignment and consistency across generated content [1].