aiqbench analysis
This preprint studies how learned policies decide which tokens to commit during parallel decoding in masked diffusion language models, a speed-accuracy trade-off that matters for inference cost. The authors isolate three design choices (labeling targets, imitation states, and observed inputs) and report gains on LLaDA-8B-Instruct with an 856K-parameter policy. Results are benchmark-specific and not yet independently replicated, so readers should weigh the claimed NFE savings against their own workloads.
What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models
arXiv:2610.31576v1 Announce Type: new Abstract: Masked diffusion language models generate tokens in parallel, but their efficiency depends on which predictions are committed. Learned unmasking policies improve the speed--accuracy trade-off, yet differ in supervision, inputs, and capacity, obscuring which choices matter. We view supervised policies as distilling a decoding procedure through a teacher and examine independently what the teacher labels as committable, on which states the policy imitates it, what the policy observes, and how much the number of function evaluations (NFE) must decrease to offset policy overhead. We find that labeling predictions matching the final output is overly permissive at low NFE, and hidden states are weak alone but can complement confidence. About a 10% NFE reduction suffices to offset policy overhead for policies up to 7.5M parameters across tested settings. We therefore train a policy on trajectory-contraction targets, which accept only commitments staying on a reference trajectory, using confidence and hidden states. On LLaDA-8B-Instruct, the resulting 856K-parameter policy achieves the best accuracy--NFE frontier and highest throughput on GSM8K, is competitive at moderate targets on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.
Title:What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models
Abstract:Masked diffusion language models generate tokens in parallel, but their efficiency depends on which predictions are committed. Learned unmasking policies improve the speed--accuracy trade-off, yet differ in supervision, inputs, and capacity, obscuring which choices matter. We view supervised policies as distilling a decoding procedure through a teacher and examine independently what the teacher labels as committable, on which states the policy imitates it, what the policy observes, and how much the number of function evaluations (NFE) must decrease to offset policy overhead. We find that labeling predictions matching the final output is overly permissive at low NFE, and hidden states are weak alone but can complement confidence. About a 10% NFE reduction suffices to offset policy overhead for policies up to 7.5M parameters across tested settings. We therefore train a policy on trajectory-contraction targets, which accept only commitments staying on a reference trajectory, using confidence and hidden states. On LLaDA-8B-Instruct, the resulting 856K-parameter policy achieves the best accuracy--NFE frontier and highest throughput on GSM8K, is competitive at moderate targets on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.
Preprint.32 pages, 11 figures
Artificial Intelligence (cs.AI)
arXiv:2610.31576 [cs.AI]
(or arXiv:2610.31576v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.31576 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 292 words
Related coverage
TIPS: A Tiny Inference-time Policy for Scaling in Diffusion Language Models
arXiv · Oct 17, 2026
Do Chess Explanations Reflect Model Decisions? Behavioral and Token-Level Tests of LLM Reasoning Faithfulness
arXiv · Sep 22, 2026
See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs
arXiv · Sep 30, 2026
Unmask the State: When Does State Adaptation Matter for Masked Diffusion Language Models
arXiv · Sep 29, 2026