frontier models
A Diffusion Model Learns Which Tokens Are Safe to Commit
A small policy trained on trajectory-contraction targets improves the accuracy-versus-compute frontier by rejecting locally confident commitments that derail the reference path.
Summary
A small policy trained on trajectory-contraction targets improves the accuracy-versus-compute frontier by rejecting locally confident commitments that derail the reference path.
Masked diffusion language models predict many positions in parallel, making the commitment policy central to both speed and accuracy. The authors separate the effects of teacher labels, imitation states, observations and policy size. Labels based only on matching the final output were too permissive at low evaluation counts, while hidden states helped when combined with confidence. An 856,000-parameter policy trained to accept commitments that stay on a reference trajectory delivered the strongest reported accuracy-versus-evaluation frontier on GSM8K and transferred substantially to another tested model. The result is limited to the evaluated models, tasks and throughput setup.
Why it matters
A small policy trained on trajectory-contraction targets improves the accuracy-versus-compute frontier by rejecting locally confident commitments that derail the reference path.
Limits and context
- Labels based only on matching the final output were too permissive at low evaluation counts, while hidden states helped when combined with confidence.
Key claims
A small policy trained on trajectory-contraction targets improves the accuracy-versus-compute frontier by rejecting locally confident commitments that derail the reference path.
Qualification: Labels based only on matching the final output were too permissive at low evaluation counts, while hidden states helped when combined with confidence.
Evidence: source-2026-10-24-004
Sources
- arXiv preprint 2610.31576, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.