TACG Decoder Uses Trajectory Signals to Improve Diffusion Language Models

Trajectory-Aware Commit Gating (TACG) is a training-free decoder for Diffusion Language Models that uses temporal implicit logits guidance. It prevents premature token commitment by anchoring identities to the base posterior and utilizing history gates to verify cross-step support.