CoDR
A training-free method for correcting early generation errors in masked diffusion language models. It identifies low-confidence tokens via confidence-drift signals, re-masks them, and regenerates, remaining compatible with multiple samplers.
Timeline (last 90 days)
CoDR (Confidence Drift Remasking) is a training-free, sampler-agnostic method for correcting early errors in diffusion language models.
The k-partition probing mechanism randomly divides committed token positions into k non-overlapping subsets, re-masking one subset at a time and performing one forward inference pass.
CoDR only re-masks and regenerates tokens that the model no longer endorses, rather than discarding and redoing the entire output.
The paper validates CoDR under a setup of two backbone models, four reasoning and code tasks, and three basic samplers.
CoDR improves average accuracy across all evaluated model-sampler combinations.
CoDR requires far fewer forward passes than previous re-masking methods.
The code for CoDR has been open-sourced on GitHub (github.com/YueWu0301/CoDR).
All Facts (7)
CoDR (Confidence Drift Remasking) is a training-free, sampler-agnostic method for correcting early errors in diffusion language models.
50%UnverifiedThe k-partition probing mechanism randomly divides committed token positions into k non-overlapping subsets, re-masking one subset at a time and performing one forward inference pass.
50%UnverifiedCoDR only re-masks and regenerates tokens that the model no longer endorses, rather than discarding and redoing the entire output.
50%UnverifiedThe paper validates CoDR under a setup of two backbone models, four reasoning and code tasks, and three basic samplers.
50%UnverifiedCoDR improves average accuracy across all evaluated model-sampler combinations.
50%UnverifiedCoDR requires far fewer forward passes than previous re-masking methods.
50%UnverifiedThe code for CoDR has been open-sourced on GitHub (github.com/YueWu0301/CoDR).
50%