What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models
The paper studies learned unmasking policies for masked diffusion language models, which generate tokens in parallel and must choose which predictions to commit. The authors examine the teacher's labeling of committable predictions, the states on which the policy imitates it, the inputs the policy observes, and how much function evaluations must drop to offset policy overhead. They report that labeling predictions that match the final output is overly permissive at low NFE, and that hidden states are weak alone but complement confidence. They train a policy on trajectory-contraction targets using confidence and hidden states, reporting an 856K-parameter policy on LLaDA-8B-Instruct with the best accuracy-NFE frontier among compared methods.
One primary source; other links are context, not independent corroboration.
- Primary sourceHugging Face Daily Papers
- Discovery linkHugging Face
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Open full article view →- Original paper+45
- Fetched source page+30
- No editor link review+0
- Original non-discussion link+5
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
The paper studies learned unmasking policies for masked diffusion language models, which generate tokens in parallel and must choose which predictions to commit. The authors examine the teacher's labeling of committable predictions, the states on which the policy imitates it, the inputs the policy observes, and how much function evaluations must drop to offset policy overhead. They report that labeling predictions that match the final output is overly permissive at low NFE, and that hidden states are weak alone but complement confidence. They train a policy on trajectory-contraction targets using confidence and hidden states, reporting an 856K-parameter policy on LLaDA-8B-Instruct with the best accuracy-NFE frontier among compared methods.
Key findings from the stored brief
- Labeling predictions that match the final output is overly permissive at low function-evaluation counts.
- Hidden states are weak alone but can complement confidence signals for unmasking decisions.
- About a 10% NFE reduction suffices to offset overhead for policies up to 7.5M parameters in tested settings.
- An 856K-parameter policy trained on trajectory-contraction targets reportedly achieves the best accuracy-NFE frontier on LLaDA-8B-Instruct.
Source claims not independently checked; the supplied text is truncated and results are reported only for the tested settings.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Related archive coverage
These relationships record a citation or topic overlap, not confirmation of the finding.
The paper page is linked. Peer review and findings have not been independently checked by AINewsIQ.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
A publication revision history is not stored for this signal. New engagement snapshots do not mean the source itself changed.
Why this may matter
Learned unmasking policies could speed up parallel decoding in diffusion language models. The study isolates which design choices (supervision, inputs, capacity) drive the speed-accuracy trade-off, offering guidance for building cheaper decoding schedulers.
Limits and open questions
Source claims not independently checked; the supplied text is truncated and results are reported only for the tested settings.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Discussion
Ask about the source, add context, or point readers to a relevant follow-up.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.