TraceLock: Learning Token Commitment for Diffusion Language Models
A new paper on arXiv introduces TraceLock, a lightweight plug-in controller that learns a token-commitment policy for diffusion large language models. Diffusion LLMs generate text by refining multiple token positions in parallel, but deciding which proposed tokens to accept at each step—termed token commitment—is a hidden control problem. Existing methods rely on hand-designed confidence rules or block-specific filters. TraceLock treats token commitment as a learnable trace-state policy, using self-supervision from future stability: a proposed token is labeled stable if it matches the final token after full decoding. The controller scores variable-length trace states to make commitment decisions. The paper is authored by researchers and published on arXiv with ID 2605.24697.
Key facts
- TraceLock is a plug-in controller for frozen diffusion language models.
- Token commitment decides which proposed tokens to transfer into the partially decoded sequence.
- Existing methods use hand-designed confidence rules or block-specific acceptance filters.
- TraceLock learns token commitment as a reusable trace-state policy.
- Self-supervision derives from future stability: a token is stable if it matches the final token after full decoding.
- The controller scores variable-length trace states at each decoding step.
- The paper is available on arXiv with ID 2605.24697.
- The work addresses a control problem in parallel decoding of diffusion LLMs.
Entities
Institutions
- arXiv