arXiv:2611.21669v1 Announce Type: new Abstract: Masked diffusion models repeatedly predict provisional tokens before a sampler decides which predictions become permanent. Existing inference-time steering methods primarily modify activations, logits, or denoising trajectories, while treating this commitment policy as a fixed decoding detail. We instead study token commitment as a diffusion-specific control surface and separate it into two decisions: which tokens to commit and how many tokens to commit at each step. For the first, we introduce counterfactual target-aware commitment, which ranks unresolved positions by a mixture of model confidence and leave-one-token-out target contribution. For the second, we introduce calibrated adaptive budgets driven by target progress and confidence. Across language and vision-language settings in the LLaDA family, multiple targets, two activation-steering methods, and an unsteered control, both mechanisms yield empirically monotonic target-quality trade-offs in every configuration we evaluate. Adaptive schedules also recover competitive operating points relative to scorer-free stochastic schedules matched in average denoising cost. Because both mechanisms require only a scalar score evaluated on a provisional completion, they apply unchanged when the target is an automatic proxy for task success rather than an output attribute: NLI entailment for dialogue summarization and CLIPScore for captioning, neither of which corresponds to a single reusable steering direction. We reproduce the behavior of both mechanisms on a second masked-diffusion family, Dream-7B, for one control target and one task-success target. These results characterize commitment--which provisional predictions become irreversible, and how rapidly--as an interpretable control surface for steering masked diffusion generation under the conditions we test.
Token Commitment as a Control Surface for Steering Masked Diffusion Generation
A November 2026 arXiv paper (2611.21669v1) treats token commitment as a control surface for steering masked diffusion models, splitting it into which tokens to commit and how many to commit per step. The authors introduce counterfactual target-aware commitment, which ranks unresolved positions by a mixture of model confidence and leave-one-token-out target contribution, and calibrated adaptive budgets driven by target progress and confidence. Across language and vision-language settings in the LLaDA family, multiple targets, two activation-steering methods and an unsteered control, both mechanisms produced empirically monotonic target-quality trade-offs in every configuration evaluated, and the behavior reproduced on the Dream-7B masked-diffusion family for one control target and one task-success target.
When steering masked diffusion generation, tune the commitment policy — which tokens to commit via confidence plus leave-one-token-out target contribution, and how many per step — since it produced monotonic target-quality trade-offs across all tested LLaDA configurations.
more in playbook: Inference and self-hostingRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.