Token Commitment as a Control Surface for Steering Masked Diffusion Generation
Organizations: Stanford University
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.
Figures & tables
| Setting | Model | Data ( ) | Activation | Target |
| Sentiment | LLaDA | Amazon ( ) | None / DiffMean / ACT | Negative sentiment |
| Subjectivity | LLaDA | Amazon ( ) | None / DiffMean / ACT | Subjectivity |
| Summarization | LLaDA | SAMSum ( ) | None | NLI entailment |
| Caption sentiment | LLaDA-V | SentiCap ( ) | None / DiffMean / ACT | Negative sentiment |
| Image alignment | LLaDA-V | SentiCap ( ) | None | CLIPScore |
| Sentiment | Dream-7B | Amazon ( ) | None / DiffMean / ACT | Negative sentiment |
| Model / target | Activation | Target | PPL | Ordering consistent |
| LLaDA / Subjectivity | None | ✓ | ||
| LLaDA / Subjectivity | DiffMean | ✓ | ||
| LLaDA / Subjectivity | ACT | ✓ | ||
| LLaDA-V / Caption sentiment | None | ✓ | ||
| LLaDA-V / Caption sentiment | DiffMean | ✓ | ||
| LLaDA-V / Caption sentiment | ACT | ✓ |
| Model / target | Activation | Target | PPL | Competitive | |
| LLaDA / Sentiment | DiffMean | ✓ | |||
| ✓ | |||||
| ✓ | |||||
| LLaDA-V / Caption sentiment | None | ✓ | |||
| ✓ | |||||
| ✓ |
Appendix figures & tables42 assets
Supplementary material from the paper’s appendix.
Appendix
| Task | Prompt template |
| Product-review continuation | Continue the following product review naturally. Keep the same topic as the original text. Write around 10 complete and long sentences. Do not add headings, labels, or commentary. Review: {prefix} |
| SAMSum summarization | Summarize the following conversation in one concise sentence. Do not add headings, labels, or commentary. Conversation: {dialogue} Summary: |
| LLaDA-V captioning | Describe the image. |
| Hyperparameter | Value |
| LLaDA model | GSAI-ML/LLaDA-8B-Instruct |
| LLaDA-V model | GSAI-ML/LLaDA-V |
| Dream model | Dream-org/Dream-v0-Instruct-7B |
| LLaDA generation length | |
| LLaDA block length | |
| LLaDA-V generation length |
| Model | Method | Target | Strength |
| LLaDA | DiffMean | Sentiment | |
| LLaDA | DiffMean | Subjectivity | |
| LLaDA | ACT | Sentiment | |
| LLaDA | ACT | Subjectivity | |
| LLaDA-V | DiffMean | Caption sentiment | |
| LLaDA-V | ACT | Caption sentiment |
| Setting | Data source | Pole sizes | Construction details |
| LLaDA DiffMean sentiment | Amazon Polarity train | positive, negative | Mean-difference vector oriented toward the negative-sentiment pole, ; text truncated to characters; tokenized to max length |
| LLaDA DiffMean subjectivity | Amazon Polarity train, scored by the subjectivity classifier | top objective, bottom objective | Same DiffMean construction as sentiment, using subjective and objective poles; vector is oriented toward the subjective target; text truncated to characters; tokenized to max length |
| LLaDA ACT sentiment | Amazon Polarity train | positive, negative | Diagonal source/target statistics from positive source to negative target; text truncated to characters; tokenized to max length |
| LLaDA ACT subjectivity | Amazon Polarity train, scored by the subjectivity classifier | top objective, bottom objective | Diagonal source/target statistics from objective source to subjective target; text truncated to characters; tokenized to max length |
| LLaDA-V DiffMean sentiment | SentiCap captions | positive, negative | Mean-difference vector oriented toward the negative-caption-sentiment pole; sign is matched to the LLaDA convention so steers toward negative sentiment; activations pooled over caption-token positions only, excluding image patch tokens |
| LLaDA-V ACT sentiment | SentiCap captions | positive, negative | Diagonal source/target statistics from positive caption-sentiment source to negative caption-sentiment target; activations pooled over caption-token positions only, excluding image patch tokens |
| Policy | Average steps | Mean s/gen | ms/ step | Runtime share (%) LLaDA / scorer / other | Overhead |
| Matched stochastic | reference | ||||
| Adaptive, | |||||
| Confidence ranking, | reference | ||||
| Counterfactual, |
| Model / target | Target | PPL | Best-of- at matched budget target / PPL |
| LLaDA / Sentiment | / | ||
| Dream-7B / Sentiment | / | ||
| LLaDA / NLI entailment | / | ||
| Dream-7B / NLI entailment | / |
| Model / target | Activation | Target | PPL | Ordering consistent target / PPL |
| LLaDA / Sentiment | None | ✓/ ✓ | ||
| LLaDA / Sentiment | DiffMean | ✓/ ✓ | ||
| LLaDA / Sentiment | ACT | ✓/ ✓ | ||
| LLaDA / Subjectivity | None | ✓/ ✓ | ||
| LLaDA / Subjectivity | DiffMean | ✓/ ✓ | ||
| LLaDA / Subjectivity | ACT | ✓/ ✓ |
| Commitment policy | Target | PPL |
| Confidence ranking ( ) | ||
| Mixed ( ) | ||
| Pure counterfactual ( ) | ||
| Random commitment order |
| Model / target | Activation | Target | PPL | Ordering consistent target / PPL |
| LLaDA / Sentiment | None | ✓/ ✓ | ||
| LLaDA / Sentiment | DiffMean | ✓/ ✓ | ||
| LLaDA / Sentiment | ACT | ✓/ ✓ | ||
| LLaDA / Subjectivity | None | ✓/ ✓ | ||
| LLaDA / Subjectivity | DiffMean | ✓/ ✓ | ||
| LLaDA / Subjectivity | ACT | ✓/ ✓ |
| Model / target | Activation | Target | PPL | Competitive |
| LLaDA / Sentiment | None | ✓/ ✓/ ✓ | ||
| LLaDA / Sentiment | DiffMean | ✓/ ✓/ ✓ | ||
| LLaDA / Sentiment | ACT | ✓/ ✓/ ✓ | ||
| LLaDA / Subjectivity | None | ✓/ ✓/ ✓ | ||
| LLaDA / Subjectivity | DiffMean | ✓/ ✓/ ✓ | ||
| LLaDA / Subjectivity | ACT | ✓/ ✓/ ✓ |
| Control | Activation | Target means (increasing ) | PPL means (increasing ) | Target order preserved | PPL order preserved |
| Commitment order | DiffMean | ||||
| Commitment order | ACT | ||||
| Commitment order | None | ||||
| Adaptive budget | DiffMean | ||||
| Adaptive budget | ACT | ||||
| Adaptive budget | None |
| Activation | |||
| DiffMean | |||
| ACT | |||
| No activation |
| Sweep | Activation | Decoding scorer (SST-2) | Held-out scorer (cardiffnlp) | Ordering preserved |
| None | ✓ | |||
| DiffMean | ✓ | |||
| ACT | ✓ | |||
| None | ✓ | |||
| DiffMean | ✓ | |||
| ACT | ✓ |
| Sweep | Activation | Decoding scorer (SST-2) | Held-out scorer (cardiffnlp) | Ordering preserved |
| None | ✓ | |||
| DiffMean | ✓ | |||
| ACT | ✓ | |||
| None | ✓ | |||
| DiffMean | ✓ | |||
| ACT | ✓ |
| Model | Sweep | Decoding scorer (nli-MiniLM2) | Held-out scorer (nli-deberta-v3-large) | Ordering preserved |
| LLaDA | ✓ | |||
| LLaDA | ✓ | |||
| Dream-7B | ✓ | |||
| Dream-7B | ✓ |
| Sweep | Decoding scorer (OpenAI ViT-B/32) | Held-out scorer (LAION-2B ViT-B/32) | Ordering preserved |
| ✓ | |||
| ✓ |
| Target | Level | Score | Generated continuation |
| Sentiment | Low | 0.000 | “The author is a master of the genre, and the book is a masterclass in comedic writing. The book is a must-read for anyone who loves a good laugh …” |
| Mid. | 0.531 | “…it didn’t work. I got the second one and it didn’t work either. I don’t know if it’s the bread machine or the seller, but I’m not going to waste my money on this particular bread machine again.” | |
| High | 1.000 | “This movie was a waste of time. The plot was weak, the acting was bad, and the special effects were terrible. Don’t waste your time or your money on this movie.” | |
| Subjectivity | Low | 0.248 | “…message through the use of visuals, sound, music, dialogue, acting, directing, writing, producing, camera, lighting, editing, sound, makeup, costumes, props, location, casting, financing, marketing, distribution, and more.” |
| Mid. | 0.665 | “However, it’s important to remember that the use of such methods is highly controversial and often associated with ethical concerns. It’s crucial to approach these topics with caution and respect …” | |
| High | 0.801 | “…ancient power and mystery. …they must use their own wits and courage to uncover the true secrets of the temple.” |
| Level | Score | Dialogue and generated summary |
| Low | 0.127 | Dialogue: Joshua: look out the window / Noah: what’s there? XD / Joshua: do it :P / Noah: oh wow / Noah: so much snow :D when did this happen lol / Joshua: :D / Noah: it’s definitely time to build a snowman! / Joshua: build an Olaf :d / Noah: ofc i will, he’s my idol! / Joshua: XD Summary: “Joshua and Noah are excited about the unexpected snow outside and plan to build a snowman, with Joshua suggesting they create an Olaf-themed snowman and Noah agreeing, as he admires fan favorite Olaf from the Wreck-It Ralph movie.” The final clause is a hallucination: Olaf is not from that film, and the dialogue says nothing about its origin. |
| Intermediate | 0.487 | Dialogue: Max: I’m so sorry Lucas. I don’t know what got into me. / Lucas: … / Lucas: I don’t know either. / Mason: that was really f***ed up Max / Max: I know. I’m so sorry :(. / Lucas: I don’t know, man. / Mason: what were you thinking?? / Max: I wasn’t. / Max: Can we please meet and talk this through? Please. / Lucas: Ok. I’ll think about it and let you know. / Max: Thanks… Summary: “Max apologizes to Lucas and Mason, expressing confusion over his actions, while Lucas and Mason also express confusion and concern, ultimately agreeing to meet and discuss the situation, with Lucas promising to think about it and let Max know when he’s ready.” The gist is correct, but Max apologizes to Lucas rather than to both, Mason expresses judgement rather than confusion, and Lucas agrees only to consider meeting. |
| High | 0.947 | Dialogue: Mike: hey have you played RDR2 already? / Jack: of course / Jack: I think I’m halfway through / Mike: ooooh / Mike: and whaddya think? / Mike: coz I’ve just started / Jack: what can I say, it’s freaking great / Jack: the storytelling, the acting, the graphics, immersiveness (is that a word? hehe) / Jack: you know, when I was a kid, I used to play with cowboy-themed lego sets / Jack: so it’s like a dream coming true for me :D / Mike: haha / Mike: totally Summary: “Jack, who has already played RDR2 and is halfway through, expresses his immense enthusiasm for the game, praising its storytelling, acting, graphics, and immersiveness, and likens it to his childhood fascination with cowboy-themed lego sets.” Every proposition is supported by the dialogue, including the specific list of praised attributes and the childhood comparison. |
| Activation | Target | PPL | Ordering consistent target / PPL |
| None | ✓/ ✓ | ||
| DiffMean | ✓/ ✓ | ||
| ACT | ✓/ ✓ |
| Activation | Target | PPL | Ordering consistent target / PPL |
| None | ✓/ ✓ | ||
| DiffMean | ✓/ ✓ | ||
| ACT | ✓/ ✓ |
| Activation | Matched-fixed Target / PPL | Target | PPL | Competitive |
| None | / | ✓/ ✓/ ✓ | ||
| DiffMean | / | ✓/ ✓/ ✓ | ||
| ACT | / | ✓/ ✓/ ✓ |
| Control | Target (increasing ) | PPL (increasing ) | Ordering consistent target / PPL |
| Commitment order, | ✓/ ✓ | ||
| Adaptive budget, | ✓/ ✓ |
| Activation | Matched-fixed Target / PPL | Target | PPL | Competitive |
| None | / | ✓/ ✓/ ✓ |
Diffusion Models
Diffusion models learn to reverse gradual data corruption, enabling flexible generation, editing, forecasting, and inference.
See all 1,266 papers