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
Figure 1: Counterfactual target-aware commitment order. Each unresolved position contributes one provisional candidate. We measure its local target contribution by deleting its decoded span, combine the normalized score drop with model confidence, and commit the highest-scoring position. With λorder=0.75 , the second-most-confident position can still win when its counterfactual contribution is sufficiently stronger.
Figure 2: Adaptive commitment budgets. Provisional predictions are always ranked by confidence. Target progress and top-candidate confidence are calibrated within the current denoising phase, mixed through λbudget , and mapped to ut∈{1,2,3} . Increasing ut commits more of the same confidence ranking and leaves fewer positions revisable.
Setting
Model
Data ( N )
Activation
Target
Sentiment
LLaDA
Amazon ( 300 )
None / DiffMean / ACT
Negative sentiment
Subjectivity
LLaDA
Amazon ( 300 )
None / DiffMean / ACT
Subjectivity
Summarization
LLaDA
SAMSum ( 300 )
None
NLI entailment
Caption sentiment
LLaDA-V
SentiCap ( 100 )
None / DiffMean / ACT
Negative sentiment
Image alignment
LLaDA-V
SentiCap ( 100 )
None
CLIPScore
Sentiment
Dream-7B
Amazon ( 300 )
None / DiffMean / ACT
Negative sentiment
Table 1: Evaluation settings. Sentiment, subjectivity and caption sentiment are control targets; NLI entailment and CLIPScore are task-success targets and use no activation intervention, isolating commitment control. The last two rows are the second-family confirmation of Appendix D , which covers one control target and one task-success target.
Figure 3: Counterfactual commitment-order trade-offs for LLaDA sentiment. We vary λorder while fixing ut≡1 , so movement along each curve changes only which provisional token is committed next.
Model / target
Activation
Target λorder=0.5→0.75→0.9→1
PPL λorder=0.5→0.75→0.9→1
Ordering consistent
LLaDA / Subjectivity
None
0.64→0.57→0.54→0.49
12.5→10.7→9.4→8.9
✓
LLaDA / Subjectivity
DiffMean
0.75→0.70→0.63→0.57
14.2→11.0→10.3→8.6
✓
LLaDA / Subjectivity
ACT
0.70→0.60→0.54→0.47
12.2→9.8→9.1→8.6
✓
LLaDA-V / Caption sentiment
None
0.84→0.47→0.33→0.32
13.5→12.3→12.0→11.8
✓
LLaDA-V / Caption sentiment
DiffMean
0.89→0.47→0.38→0.33
15.0→12.1→11.6→11.3
✓
LLaDA-V / Caption sentiment
ACT
0.57→0.49→0.45→0.42
11.5→11.1→11.1→11.0
✓
Table 2: Generalization of counterfactual commitment order across control targets and across model families. We report the full target-score and PPL sequences as λorder moves from stronger counterfactual weighting ( 0.5 ) toward confidence-only commitment ( 1 ). ✓ indicates that both sequences are non-increasing across the four tested weights.
Figure 4: Adaptive commitment budgets for LLaDA sentiment with ACT. Varying λbudget with confidence ranking fixed yields an empirically ordered target–quality sweep.
Figure 5: Adaptive budgets versus a scorer-free stochastic fixed schedule matched in average denoising steps, for LLaDA sentiment with ACT. Fixed- 1/2/3 are shown as contextual reference points. This matched comparison isolates state-dependent adaptation from average denoising cost.
Model / target
Activation
λbudget
Δ Target
Δ PPL
Competitive
LLaDA / Sentiment
DiffMean
0
+0.07
+4.2
✓
0.5
−0.03
−3.0
✓
1
−0.07
−4.2
✓
LLaDA-V / Caption sentiment
None
0
+0.24
+2.5
✓
0.5
−0.01
−1.1
✓
1
−0.15
−2.9
✓
Table 3: Adaptive schedules versus compute-matched stochastic fixed schedules, for four representative settings; all fifteen are in Appendices C.3 and D . Deltas are adaptive minus matched stochastic, so positive Δ Target favors adaptation and negative Δ PPL indicates better fluency. “Competitive” means the adaptive point is not jointly dominated in target and PPL at matched average denoising cost. The Dream-7B λbudget=0 row improves both metrics simultaneously.
Figure 6: Counterfactual commitment order for task-success objectives. Left: mean NLI entailment probability versus PPL on SAMSum. Right: mean CLIPScore versus PPL on SentiCap, where CLIPScore is nonnegative cosine similarity scaled by 100 . Both use ut≡1 and no activation steering.
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.
Appendix
Table 4: Prompt templates used in the experiments.
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
60
LLaDA block length
60
LLaDA-V generation length
35
Appendix
Table 5: Generation, scoring, and commitment hyperparameters shared across experiments. The held-out scorers are used only for the post-hoc checks in Appendix C.5 and never during decoding. Activation-steering strengths are reported separately in Table 6 .
Model
Method
Target
Strength
LLaDA
DiffMean
Sentiment
α=3.0
LLaDA
DiffMean
Subjectivity
α=2.0
LLaDA
ACT
Sentiment
β=0.35
LLaDA
ACT
Subjectivity
β=0.25
LLaDA-V
DiffMean
Caption sentiment
α=3.0
LLaDA-V
ACT
Caption sentiment
β=0.5
Appendix
Table 6: Activation-steering strengths used in the experiments. Each is held at a single value throughout, so that the commitment weight is the only variable along a sweep. By convention, positive α or β always steers toward the paper’s stated target direction: negative sentiment, subjectivity, or negative caption sentiment. DiffMean applies hℓ←hℓ+αvℓ , while ACT applies hℓ←hℓ+β(Tℓ(hℓ)−hℓ) .
Setting
Data source
Pole sizes
Construction details
LLaDA DiffMean sentiment
Amazon Polarity train
5000 positive, 5000 negative
Mean-difference vector oriented toward the negative-sentiment pole, vℓ=μℓneg−μℓpos ; text truncated to 4000 characters; tokenized to max length 256
LLaDA DiffMean subjectivity
Amazon Polarity train, scored by the subjectivity classifier
top 5000 objective, bottom 5000 objective
Same DiffMean construction as sentiment, using subjective and objective poles; vector is oriented toward the subjective target; text truncated to 4000 characters; tokenized to max length 256
LLaDA ACT sentiment
Amazon Polarity train
5000 positive, 5000 negative
Diagonal source/target statistics from positive source to negative target; text truncated to 4000 characters; tokenized to max length 256
LLaDA ACT subjectivity
Amazon Polarity train, scored by the subjectivity classifier
top 5000 objective, bottom 5000 objective
Diagonal source/target statistics from objective source to subjective target; text truncated to 4000 characters; tokenized to max length 256
LLaDA-V DiffMean sentiment
SentiCap captions
1000 positive, 1000 negative
Mean-difference vector oriented toward the negative-caption-sentiment pole; sign is matched to the LLaDA convention so +α steers toward negative sentiment; activations pooled over caption-token positions only, excluding image patch tokens
LLaDA-V ACT sentiment
SentiCap captions
1000 positive, 1000 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
Appendix
Table 7: Activation-vector and ACT-statistics construction. These hyperparameters determine what each activation intervention steers toward.
Policy
Average steps
Mean s/gen
ms/ step
Runtime share (%) LLaDA / scorer / other
Overhead
Matched stochastic
33.280
0.6584
19.78
98.5/0.0/1.5
reference
Adaptive, λbudget=0.5
33.273
0.7141
21.46
90.5/6.2/3.3
+8.5%
Confidence ranking, ut=1
60.000
1.1871
19.79
98.5/0.0/1.5
reference
Counterfactual, λorder=0.5
60.000
1.7676
29.46
66.4/18.4/15.2
+48.9%
Appendix
Table 8: Representative runtime overhead. “Other” includes provisional-text decoding, deletion-variant construction, calibration and normalization, and commitment-policy operations. Overhead is measured relative to the scorer-free baseline with either matched average denoising cost or the same fixed commitment budget.
Model / target
Target λorder=0.5→0.75→0.9→1
PPL λorder=0.5→0.75→0.9→1
Best-of- N at matched budget target / PPL
LLaDA / Sentiment
0.63→0.53→0.48→0.35
10.6→10.5→9.6→9.5
0.483 / 10.92
Dream-7B / Sentiment
0.71→0.52→0.41→0.30
8.0→7.8→7.4→7.0
0.515 / 7.95
LLaDA / NLI entailment
0.54→0.43→0.34→0.31
26.6→25.1→23.8→22.5
0.404 / 25.51
Dream-7B / NLI entailment
0.60→0.47→0.39→0.34
27.3→25.0→23.5→22.4
0.470 / 26.12
Appendix
Table 9: Commitment-order curves against best-of- N at a matched generation budget. The first two columns reproduce the commitment-order sequences of Table 10 for the four settings evaluated without activation steering; the last column is the best-of- N operating point at the expected generation count matching that setting’s counterfactual-ordering overhead.
Model / target
Activation
Target λorder=0.5→0.75→0.9→1
PPL λorder=0.5→0.75→0.9→1
Ordering consistent target / PPL
LLaDA / Sentiment
None
0.63→0.53→0.48→0.35
10.6→10.5→9.6→9.5
✓/ ✓
LLaDA / Sentiment
DiffMean
0.80→0.62→0.56→0.50
11.5→11.0→9.6→8.9
✓/ ✓
LLaDA / Sentiment
ACT
0.75→0.69→0.62→0.51
11.2→11.1→10.1→9.2
✓/ ✓
LLaDA / Subjectivity
None
0.64→0.57→0.54→0.49
12.5→10.7→9.4→8.9
✓/ ✓
LLaDA / Subjectivity
DiffMean
0.75→0.70→0.63→0.57
14.2→11.0→10.3→8.6
✓/ ✓
LLaDA / Subjectivity
ACT
0.70→0.60→0.54→0.47
12.2→9.8→9.1→8.6
✓/ ✓
Appendix
Table 10: Full counterfactual commitment-order sequences. Target scores and PPL are reported as λorder moves from stronger counterfactual weighting ( 0.5 ) toward confidence-only commitment ( 1 ). The final column indicates whether the target and PPL sequences are non-increasing over the tested commitment weights. Full-precision Dream-7B values are in Tables 22 and 25 .
Figure 7: Counterfactual commitment-order trade-off plot for LLaDA subjectivity control.
Table 11: Extreme commitment-order diagnostics for LLaDA sentiment without activation steering. All four policies use ut≡1 and therefore the same 60 denoising steps. Random ordering incurs the largest perplexity cost of any policy shown while barely moving the target; the intermediate rows are reproduced from Table 10 for reference.
Model / target
Activation
Target λbudget=0→0.5→1
PPL λbudget=0→0.5→1
Ordering consistent target / PPL
LLaDA / Sentiment
None
0.51→0.39→0.33
18.7→15.6→12.5
✓/ ✓
LLaDA / Sentiment
DiffMean
0.65→0.55→0.51
21.3→14.1→12.9
✓/ ✓
LLaDA / Sentiment
ACT
0.65→0.57→0.52
21.1→15.6→13.0
✓/ ✓
LLaDA / Subjectivity
None
0.61→0.55→0.51
18.4→13.2→11.6
✓/ ✓
LLaDA / Subjectivity
DiffMean
0.61→0.60→0.58
19.7→13.8→10.5
✓/ ✓
LLaDA / Subjectivity
ACT
0.62→0.57→0.52
17.5→13.3→10.5
✓/ ✓
Appendix
Table 12: Full adaptive-budget sequences. Target scores and PPL are reported for λbudget∈{0,0.5,1} . The scoring map is fixed to confidence throughout, so movement across each sequence changes only the adaptive commitment budget. Full-precision Dream-7B values are in Tables 23 and 25 .
Figure 9: Adaptive-budget sweeps for LLaDA sentiment control under no activation steering, DiffMean, and ACT.
Figure 10: Adaptive-budget sweeps for LLaDA subjectivity control under no activation steering, DiffMean, and ACT.
Figure 11: Adaptive-budget sweeps for LLaDA-V caption sentiment under no activation steering, DiffMean, and ACT. The CLIPScore panel is the image-grounding check described above.
Figure 12: Adaptive-budget sweep for NLI entailment.
Figure 13: Adaptive-budget sweep for CLIPScore image–text alignment.
Model / target
Activation
Δ Target λbudget=0→0.5→1
Δ PPL λbudget=0→0.5→1
Competitive 0/0.5/1
LLaDA / Sentiment
None
+0.14→+0.02→−0.04
+1.6→−1.5→−4.6
✓/ ✓/ ✓
LLaDA / Sentiment
DiffMean
+0.07→−0.03→−0.07
+4.2→−3.0→−4.2
✓/ ✓/ ✓
LLaDA / Sentiment
ACT
+0.04→−0.04→−0.09
+4.9→−0.6→−3.2
✓/ ✓/ ✓
LLaDA / Subjectivity
None
+0.03→−0.03→−0.07
+2.8→−2.4→−4.0
✓/ ✓/ ✓
LLaDA / Subjectivity
DiffMean
+0.02→+0.02→−0.01
+3.3→−2.6→−6.0
✓/ ✓/ ✓
LLaDA / Subjectivity
ACT
+0.05→+0.00→−0.04
+4.1→−0.1→−2.9
✓/ ✓/ ✓
Appendix
Table 13: Full compute-matched stochastic comparisons. For each setting, entries from left to right correspond to λbudget∈{0,0.5,1} . Each adaptive policy is compared with the matched scorer-free, non-adaptive stochastic schedule for that setting (Appendix A.4 ). Deltas are adaptive minus matched stochastic: positive Δ Target favors adaptation, while negative Δ PPL indicates lower perplexity. “Competitive” indicates that the adaptive point is not jointly dominated in target and PPL by its compute-matched stochastic baseline. Two Dream-7B sentiment entries—DiffMean at λbudget=0 and no activation at λbudget=0.5 —improve both metrics simultaneously.
Figure 14: Adaptive versus compute-matched stochastic fixed schedules for LLaDA sentiment under no activation steering, DiffMean, and ACT. The matched baseline is scorer-free, non-adaptive, confidence-ranked, and matched in average denoising steps.
Figure 15: Adaptive versus compute-matched stochastic fixed schedules for LLaDA subjectivity under no activation steering, DiffMean, and ACT.
Figure 16: Adaptive versus compute-matched stochastic fixed schedules for LLaDA-V caption sentiment under no activation steering, DiffMean, and ACT.
Figure 17: Adaptive versus compute-matched stochastic fixed schedules for the NLI entailment and CLIPScore task-success objectives. Both settings use no activation steering.
Control
Activation
Target means (increasing λ )
PPL means (increasing λ )
Target order preserved
PPL order preserved
Commitment order
DiffMean
0.799→0.615→0.562→0.500
11.54→11.05→9.60→8.88
88.1%
83.7%
Commitment order
ACT
0.751→0.695→0.625→0.509
11.23→11.13→10.08→9.18
91.4%
85.0%
Commitment order
None
0.630→0.533→0.485→0.350
10.63→10.47→9.60→9.46
91.0%
77.5%
Adaptive budget
DiffMean
0.650→0.549→0.510
21.32→14.08→12.92
86.7%
95.9%
Adaptive budget
ACT
0.653→0.574→0.516
21.07→15.60→13.02
91.0%
96.7%
Adaptive budget
None
0.508→0.390→0.332
18.68→15.58→12.48
95.3%
97.4%
Appendix
Table 14: Ordering stability for LLaDA sentiment. Means are shown at full precision as λ increases across the tested weights ( λorder=0.5→0.75→0.9→1 for commitment order, λbudget=0→0.5→1 for adaptive budget), and correspond to the rounded entries in Tables 10 and 12 . The last two columns give the fraction of 10,000 resamples of the evaluation set in which the observed ordering is reproduced.
Activation
λbudget=0
λbudget=0.5
λbudget=1
DiffMean
94.7%
99.9%
100.0%
ACT
94.7%
92.4%
99.8%
No activation
99.9%
91.1%
100.0%
Appendix
Table 15: Stability of the compute-matched comparison, LLaDA sentiment. Each entry is the fraction of 10,000 resamples in which the adaptive policy is not jointly dominated by its compute-matched stochastic schedule, that is, in which it is not simultaneously worse in target and worse in perplexity. The quantity is a floor rather than a measure of improvement: it can fall toward zero when the adaptive policy is worse on both metrics at once, but a high value indicates only that the adaptive point lies on a sensible trade-off curve relative to the scorer-free control, not that adaptation buys a particular amount.
Sweep
Activation
Decoding scorer (SST-2)
Held-out scorer (cardiffnlp)
Ordering preserved
λorder=0.5→0.75→0.9→1
None
0.63→0.53→0.48→0.35
0.40→0.38→0.33→0.26
✓
DiffMean
0.80→0.62→0.56→0.50
0.57→0.39→0.38→0.36
✓
ACT
0.75→0.69→0.62→0.51
0.55→0.52→0.47→0.41
✓
λbudget=0→0.5→1
None
0.51→0.39→0.33
0.36→0.28→0.24
✓
DiffMean
0.65→0.55→0.51
0.42→0.41→0.38
✓
ACT
0.65→0.57→0.52
0.47→0.45→0.38
✓
Appendix
Table 16: Held-out-scorer check, LLaDA sentiment. The ordering induced by λorder and λbudget is preserved in all six sweeps.
Sweep
Activation
Decoding scorer (SST-2)
Held-out scorer (cardiffnlp)
Ordering preserved
λorder=0.5→0.75→0.9→1
None
0.71→0.52→0.41→0.30
0.40→0.32→0.27→0.22
✓
DiffMean
0.80→0.64→0.54→0.44
0.59→0.48→0.43→0.37
✓
ACT
0.76→0.65→0.50→0.42
0.47→0.41→0.39→0.34
✓
λbudget=0→0.5→1
None
0.50→0.45→0.43
0.33→0.30→0.27
✓
DiffMean
0.69→0.64→0.62
0.54→0.51→0.49
✓
ACT
0.64→0.59→0.52
0.47→0.44→0.39
✓
Appendix
Table 17: Held-out-scorer check, Dream-7B sentiment. The same check applied to the second model family. The ordering is preserved in all six sweeps.
Model
Sweep
Decoding scorer (nli-MiniLM2)
Held-out scorer (nli-deberta-v3-large)
Ordering preserved
LLaDA
λorder=0.5→0.75→0.9→1
0.54→0.43→0.34→0.31
0.91→0.88→0.85→0.84
✓
LLaDA
λbudget=0→0.5→1
0.40→0.35→0.31
0.88→0.85→0.83
✓
Dream-7B
λorder=0.5→0.75→0.9→1
0.60→0.47→0.39→0.34
0.96→0.91→0.87→0.85
✓
Dream-7B
λbudget=0→0.5→1
0.43→0.34→0.26
0.89→0.85→0.82
✓
Appendix
Table 18: Held-out-scorer check, NLI entailment on both model families. The ordering induced by both commitment weights is preserved in all four sweeps. The held-out model operates over a compressed range, so the sweeps span a narrower interval than under the decoding scorer while retaining the same direction.
Sweep
Decoding scorer (OpenAI ViT-B/32)
Held-out scorer (LAION-2B ViT-B/32)
Ordering preserved
λorder=0.5→0.75→0.9→1
30.9→26.8→25.7→25.5
28.1→22.9→21.0→20.3
✓
λbudget=0→0.5→1
29.4→28.6→28.0
27.1→25.7→24.9
✓
Appendix
Table 19: Held-out-scorer check, LLaDA-V CLIPScore. Both commitment weights induce the same ordering under a CLIP model pretrained on a different corpus. Absolute values are lower throughout because the two models are calibrated differently; the held-out sweep also spans a wider interval than the decoding scorer’s.
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.”
Appendix
Table 20: Qualitative examples for the text control objectives. Scores denote the corresponding frozen target-class probability: negative sentiment or subjectivity.
Figure 18: Qualitative caption-sentiment examples. Images and generated captions are shown at increasing frozen negative-sentiment scores.
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.
Appendix
Table 21: Qualitative NLI-entailment examples. Higher scores correspond to stronger agreement between the generated summary and its source dialogue. Profanity in the source dialogue is masked.
Table 22: Counterfactual commitment-order sequences, Dream-7B sentiment. Both the target score and PPL are non-increasing across the four tested weights under all three activation conditions.
Figure 20: Counterfactual commitment-order trade-offs for Dream-7B sentiment. As in Fig. 3 , ut≡1 is fixed, so movement along each curve changes only which provisional token is committed next.
Activation
Target λbudget=0→0.5→1
PPL λbudget=0→0.5→1
Ordering consistent target / PPL
None
0.501→0.453→0.432
13.76→11.45→9.34
✓/ ✓
DiffMean
0.685→0.638→0.621
11.33→9.66→8.44
✓/ ✓
ACT
0.645→0.588→0.523
12.88→10.46→8.70
✓/ ✓
Appendix
Table 23: Adaptive-budget sequences, Dream-7B sentiment. The scoring map is fixed to confidence, so movement across each sequence changes only the commitment budget. Target-progress-driven budgeting ( λbudget=0 ) gives the strongest target and confidence-driven budgeting ( λbudget=1 ) the lowest PPL.
Figure 21: Adaptive-budget sweeps for Dream-7B sentiment under no activation steering, DiffMean, and ACT.
Activation
Matched-fixed Target / PPL
Δ Target λbudget=0→0.5→1
Δ PPL λbudget=0→0.5→1
Competitive 0/0.5/1
None
0.441 / 13.68
+0.06→+0.01→−0.01
+0.1→−2.2→−4.3
✓/ ✓/ ✓
DiffMean
0.655 / 12.15
+0.03→−0.02→−0.03
−0.8→−2.5→−3.7
✓/ ✓/ ✓
ACT
0.611 / 12.71
+0.03→−0.02→−0.09
+0.2→−2.3→−4.0
✓/ ✓/ ✓
Appendix
Table 24: Adaptive versus compute-matched stochastic fixed schedules, Dream-7B sentiment. Each adaptive configuration is compared with the same scorer-free, non-adaptive stochastic schedule for that activation condition, constructed as in Appendix A.4 and matched in average denoising steps. Deltas are adaptive minus matched stochastic. “Competitive” has the same meaning as in Table 13 . Two entries are strict improvements on both metrics at once—DiffMean at λbudget=0 ( +0.03 target, −0.8 PPL) and no activation at λbudget=0.5 ( +0.01 target, −2.2 PPL).
Figure 22: Adaptive versus compute-matched stochastic fixed schedules for Dream-7B sentiment under no activation steering, DiffMean, and ACT. Fixed- 1/2/3 are shown as contextual anchors; the matched stochastic point is the controlled comparison.
Control
Target (increasing λ )
PPL (increasing λ )
Ordering consistent target / PPL
Commitment order, λorder=0.5→0.75→0.9→1
0.597→0.469→0.393→0.337
27.35→25.01→23.50→22.40
✓/ ✓
Adaptive budget, λbudget=0→0.5→1
0.430→0.336→0.263
33.28→28.60→24.83
✓/ ✓
Appendix
Table 25: Commitment-order and adaptive-budget sequences, Dream-7B NLI entailment. Both mechanisms behave as on LLaDA: the target score and PPL are non-increasing as the commitment weight moves toward its confidence-driven end. Commitment order holds ut≡1 , so its target movement occurs at a fixed 60 denoising steps.
Activation
Matched-fixed Target / PPL
Δ Target λbudget=0→0.5→1
Δ PPL λbudget=0→0.5→1
Competitive 0/0.5/1
None
0.343 / 31.50
+0.09→−0.01→−0.08
+1.8→−2.9→−6.7
✓/ ✓/ ✓
Appendix
Table 26: Adaptive versus compute-matched stochastic fixed schedule, Dream-7B NLI entailment. Deltas are adaptive minus matched stochastic, at matched average denoising cost. All three adaptive configurations are competitive.
Figure 23: Counterfactual commitment-order trade-off for Dream-7B NLI entailment. Mean entailment probability versus PPL on SAMSum, with ut≡1 and no activation steering.
Figure 24: Adaptive-budget sweep for Dream-7B NLI entailment.
Figure 25: Adaptive versus compute-matched stochastic fixed schedule for Dream-7B NLI entailment. The matched baseline is scorer-free, non-adaptive, confidence-ranked, and matched in average denoising steps.
Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The model is trained on randomly masked sequences, whereas inference follows a trajectory shaped by the model's own predictions. Additionally, each step has no access to what the previous one computed. Recent methods narrow these limitations from separate angles, leaving open how these choices interact. We introduce PUMBA, a unified framework for trajectory-aware training that trains the denoiser on consecutive steps of policy-induced trajectories, passes information between steps, and optimizes them jointly by backpropagation through time. A controlled study of this design space shows that i) exact train--inference alignment fails due to local overfitting, whereas a looser alignment still brings training masks closer to those seen at inference; ii) passing continuous information outperforms discrete gradient estimators through the commitment at each step; and iii) performance improves as backpropagation through time spans more steps, which we support theoretically. Combined, these components match the best checkpoint of a same-size autoregressive model. Building on these findings, we scale PUMBA to supervised fine-tuning of LLaDA-8B, where it improves the trade-off between performance and number of function evaluations (NFEs) in both full-canvas and block diffusion generation. At matched performance, it needs up to 22% fewer NFEs than standard fine-tuning with twice the budget in full-canvas generation, and up to 26% fewer than standard fine-tuning for the same number of steps in block diffusion.
Diffusion large language models promise faster generation by refining many token positions in parallel, but this parallelism introduces a hidden control problem: which proposed tokens should be transferred into the partially decoded sequence at each step? We refer to this decision as token commitment. Existing frozen-generator decoders largely rely on hand-designed confidence rules or block-specific acceptance filters. We argue that token commitment can instead be learned as a reusable trace-state policy. We introduce TraceLock, a lightweight plug-in controller that instantiates this policy for a frozen diffusion language model. Since oracle commitment times are unavailable, TraceLock derives self-supervision from future stability: at decoding step t, a proposed token for position i is labeled stable if it matches the final token at position i after the full decoding trace completes. The controller scores variable-length trace states and decides which active token proposals should be committed to the partially decoded sequence. Once trained for a given frozen backbone, the controller can be deployed across local-window widths, generation lengths, and step budgets without retraining or per-setting calibration. Experiments on question answering, mathematical reasoning, and code generation show that TraceLock improves the quality-step tradeoff over heuristic and learned baselines, with particularly stable behavior under cross-setting deployment. Diagnostic analyses show that its decisions are not reducible to scalar confidence, suggesting that frozen diffusion language models expose a learnable space of commitment trajectories beyond confidence-based decoding. Code is available at https://github.com/BobSun98/TraceLock.
Bohang Sun, Max Zhu, Francesco Caso +5
Department of Computer Science and Technology, University of Cambridge · Department of Engineering Science, University of Oxford
Masked diffusion language models generate tokens in parallel, but their efficiency depends on which predictions are committed. Learned unmasking policies improve the speed--accuracy trade-off, yet differ in supervision, inputs, and capacity, obscuring which choices matter. We view supervised policies as distilling a decoding procedure through a teacher and examine independently what the teacher labels as committable, on which states the policy imitates it, what the policy observes, and how much the number of function evaluations (NFE) must decrease to offset policy overhead. We find that labeling predictions matching the final output is overly permissive at low NFE, and hidden states are weak alone but can complement confidence. About a 10% NFE reduction suffices to offset policy overhead for policies up to 7.5M parameters across tested settings. We therefore train a policy on trajectory-contraction targets, which accept only commitments staying on a reference trajectory, using confidence and hidden states. On LLaDA-8B-Instruct, the resulting 856K-parameter policy achieves the best accuracy--NFE frontier and highest throughput on GSM8K, is competitive at moderate targets on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.
Jinya Sakurai, Xun Xu, Jaehong Yoon
NTU Singapore · Institute for Advanced Intelligence and Computing (IAIC), A*STAR · Centre for Frontier AI Research (CFAR), A*STAR