Remask, Don't Replace: Token-to-Mask Refinement in Masked Diffusion Language Models
A training-free edit rule boosts math accuracy by 5.92 points by resetting bad tokens to 'mask' instead of replacing them.
Researcher Lin Yao has introduced a novel, training-free method called Token-to-Mask (T2M) refinement to address a critical flaw in how masked diffusion language models like LLaDA2.1 correct their own mistakes. These models traditionally use a Token-to-Token (T2T) rule, which overwrites a low-confidence token with a new prediction. Yao identified three structural failure modes in T2T: it fails when no single alternative is confident, it computes replacements within an already erroneous context, and its training on uniform noise doesn't match the coherent errors models actually make.
T2M offers a simpler, more effective alternative. When a token is flagged as suspect by one of three new detection heuristics, the system doesn't replace it. Instead, it resets the position back to a mask token. This allows the next denoising step to re-predict the value from a corrected, in-distribution context, which Yao argues is a better conditioning signal than an error. The method requires no new parameters or retraining, modifying only the inference-time editing rule.
The impact is significant, particularly on tasks requiring exact token-level precision. Across eight benchmarks, T2M improved accuracy, with its largest gain being a +5.92 point boost on the CMATH mathematical reasoning dataset. Analysis showed that 79.9% of baseline errors were 'last-mile corruption'—cases where the model's internal reasoning was correct, but it produced a garbled final token. T2M successfully repaired 41.3% of these frustrating failure cases, demonstrating its power to salvage otherwise correct reasoning.
- Proposes Token-to-Mask (T2M), a training-free edit rule that resets suspect tokens to a mask state instead of replacing them, improving prediction context.
- Achieved a +5.92 point accuracy gain on CMATH, repairing 41.3% of 'last-mile corruption' errors where correct reasoning was followed by a wrong final token.
- Addresses three key flaws in standard Token-to-Token editing: trigger failure, error-prone context, and mismatched training data for error correction.
Why It Matters
This simple, cost-free tweak could significantly boost the reliability of AI models for coding, math, and any task where the final token must be perfect.