Machine translation with large language models requires both high translation quality and low generation latency. Accelerating translation while preserving quality is therefore essential for practical deployment. Diffusion large language models (dLLMs) can generate tokens at arbitrary target positions through iterative denoising, offering a new paradigm for parallel translation beyond sequential autoregressive decoding. However, existing methods still incur redundant computation: fixed target generation lengths waste computation on unused positions, while confidence-only commitment delays medium-confidence tokens whose predictions have largely stabilized. To address these issues, we propose TRAC, a training-free adaptive decoding method for diffusion machine translation. TRAC adapts the target generation length using source length and translation direction, with a conservative fallback to the maximum length for empty or unusually short outputs. It preserves high-confidence commitment while identifying temporally stable medium-confidence tokens for early commitment using cross-step top-1 persistence and approximate Jensen-Shannon divergence over Top-K distributions. On bidirectional Chinese-English translation across FLORES-200, WMT23, and NTREX-128, TRAC reduces model forward calls by 24.9% to 36.8% and achieves sentence-level generation speedups of 1.91 to 2.19 times over Fast-dLLM, with only a 0.02-point decrease in macro-average COMET. These results demonstrate that adaptive generation lengths and stability-aware token commitment improve the efficiency of diffusion machine translation.
Figures & tables
Figure 1: Spatial and temporal redundancy in diffusion machine translation. (a) A stepwise Fast-dLLM decoding trajectory: each row represents one model forward pass, which predicts all target positions simultaneously. Color intensity encodes prediction confidence. (b) The fixed 128-token canvas substantially exceeds the effective output length across translation settings. (c) A trajectory audit of 400 FLORES samples reveals many positions that satisfy temporal convergence conditions but remain masked and participate in subsequent forward passes.
Figure 2: Temporal analysis of medium-confidence predictions under Fast-dLLM. (a) Most observations are already stable. (b) Already-stable observations exhibit greater top-1 persistence. (c) They also exhibit lower cross-step sparse JSD. Labels are assigned retrospectively using the complete baseline trajectory and are used only for analysis. Error bars denote 95% confidence intervals from 2,000 sentence-level bootstrap resamples stratified by translation direction.
Figure 3: Overview of TRAC. Given a source sentence, TRAC first allocates a translation-aware adaptive canvas according to source length and translation direction, then performs iterative denoising with dual-channel token commitment to commit both high-confidence and temporally stable tokens.
Model / Decoder
Metric
FLORES
WMT23
NTREX
Macro Avg.
En → Zh
Zh → En
En → Zh
Zh → En
En → Zh
Zh → En
Qwen2.5-7B
COMET ↑
87.62
86.80
85.00
79.74
84.93
84.19
84.71
Latency (s) ↓
0.646
0.706
0.517
0.805
0.671
0.716
0.677
Llama-3-8B
COMET ↑
82.08
85.57
79.74
78.62
78.98
82.70
81.28
Latency (s) ↓
0.892
0.830
0.681
0.896
0.914
0.835
0.841
Vanilla LLaDA
COMET ↑
86.54
84.87
84.34
77.95
83.01
83.05
83.29
Table 1: Main results across six dataset–direction settings. Each cell reports COMET (top) and sentence-level latency in seconds (bottom). For TRAC , parentheses indicate speedup over Fast-dLLM.
Figure 4: Macro-averaged quality and latency across the six dataset–direction settings.
Method
COMET ↑
Latency (s) ↓
NFE ↓
Fast-dLLM
82.74
1.145
22.8
Canvas only
83.52
0.711
20.5
Temporal only
82.35
0.922
18.4
TRAC
82.72
0.573
16.0
Table 2: Macro-average results for ablations of the core modules.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Data
Method
COMET
chrF++
BLEU
Lat.
NFE
FLORES
Qwen
87.62
26.76
41.18
0.646
–
Llama
82.08
19.26
28.50
0.892
–
Vanilla
86.54
26.98
39.86
9.235
128.00
Fast-dLLM
85.97
25.53
37.51
1.232
24.42
TRAC
85.21
25.05
37.68
0.616
17.46
WMT23
Qwen
85.00
28.93
41.64
0.517
–
Appendix
Table 3: Detailed main results. Latency is measured in seconds per sentence.
Figure 5: Per-setting adaptive canvas statistics. Left: mean initial canvas length for Canvas only and TRAC; the dashed line indicates a fixed 128-token canvas. Right: fallback rates for the two configurations.
Acceptance rule
Stable precision (95% CI)
Stable recall
Coverage
Earlier NFE
Medium confidence only
76.79 [75.36, 78.22]
100.00
100.00
1.94
+ top-1 persistence
79.60 [77.70, 81.57]
51.90
50.07
1.91
+ persistence + sparse JSD
79.82 [77.78, 81.93]
46.59
44.82
1.96
Appendix
Table 5: Temporal commitment audit of frozen Fast-dLLM trajectories (400 FLORES sentences, 200 per direction). A positive label indicates that the current top-1 prediction remains unchanged in subsequent masked observations and matches the token ultimately committed by Fast-dLLM. Confidence intervals use 2,000 sentence-level bootstrap resamples stratified by translation direction.
En → Zh
Zh → En
Allocation policy
COMET ↑
Lat. ↓
Δ COMET
COMET ↑
Lat. ↓
Δ COMET
Fixed-32
82.69
0.809
-1.52
75.74
0.673
-4.38
Fixed-64
83.99
1.058
-0.22
79.31
0.887
-0.81
Fixed-96
84.56
1.424
+0.35
79.79
1.248
-0.33
Fixed-128
84.21
1.683
+0.00
80.12
1.481
+0.00
Adaptive canvas (ours)
85.08
1.048
+0.87
80.96
1.026
+0.84
Appendix
Table 6: Canvas policies at confidence 0.8. Δ COMET is relative to fixed-128.
En → Zh
Zh → En
Commitment policy
COMET ↑
Lat. ↓
NFE ↓
COMET ↑
Lat. ↓
NFE ↓
Confidence-only, τ=0.5
81.90
0.656
18.65
78.05
0.523
13.97
Confidence-only, τ=0.6
83.57
0.800
23.00
79.89
0.706
18.93
Confidence-only, τ=0.7
84.50
0.961
27.05
80.96
0.859
23.66
Confidence-only, τ=0.8
85.08
1.048
30.50
80.96
1.026
28.51
Confidence-only, τ=0.9
85.15
1.152
33.76
80.89
1.243
34.15
Appendix
Table 7: Commitment policies under the same adaptive canvas. TRAC fixes high confidence at 0.8; 0.5 is the original lower threshold.
En → Zh
Zh → En
τlow
Rule
COMET
Lat.
Δ COMET [95% CI]
COMET
Lat.
Δ COMET [95% CI]
0.5
Persistence
84.27
0.863
–
80.05
0.767
–
0.5
+ JSD (TRAC)
84.35
0.901
+0.084[−0.126,+0.319]
80.23
0.817
+0.174[−0.086,+0.416]
0.6
Persistence
84.90
0.963
–
80.79
0.905
–
0.6
+ JSD (TRAC)
85.03
0.996
+0.125[+0.007,+0.294]
80.89
0.939
+0.096[−0.040,+0.329]
0.7
Persistence
85.01
1.028
–
80.99
0.985
–
Appendix
Table 8: Matched JSD ablation. Differences and 2,000-resample paired intervals compare JSD with persistence; – indicates an uncomputed interval.
En → Zh
Zh → En
Adaptive decoding policy
COMET ↑
Lat. ↓
NFE ↓
COMET ↑
Lat. ↓
NFE ↓
LESS + cache, τ=0.5
83.47
0.989
27.98
79.67
0.935
25.28
LESS + cache, τ=0.6
84.76
1.109
31.04
80.47
1.114
30.22
LESS + cache, τ=0.75
85.15
1.208
34.56
80.87
1.292
35.17
LESS + cache, τ=0.8
85.13
1.273
35.53
80.95
1.347
36.91
EV, aligned candidates
84.82
1.130
31.50
80.99
1.114
30.00
Appendix
Table 9: All adaptive-cache LESS and EV selector settings, with common reference policies. Gray: TRAC. Bold: column best at displayed precision, including ties. Latency includes probes and retries.
Diffusion language model decoding generates text through iterative parallel refinement, balancing fewer denoising steps with accuracy, coherence, and control.
Machine translation tests masked diffusion language models (dLLMs) because every source token must be rendered faithfully, while fixed canvas decoding must choose target length before denoising. Existing masked diffusion decoding work mainly studies token unmasking order, leaving this length decision under-explored despite its direct effect on coverage and redundancy. We introduce Entropy-Valley (EV), a training-free length selector that scores candidate target canvases by mean predictive entropy from all-mask forward passes and selects the canvas the backbone is most prepared to fill. Relative to a baseline using training corpus length statistics, EV recovers 64.9%, 65.3%, and 33.0% of the COMET-22 gain from reference target lengths on En→Zh, Zh→En, and En→De. Our diagnostics show that denoising-friendly lengths need not match reference lengths. Evaluation by three translation experts supports the En↔Zh adequacy gains, with stronger evidence on Zh→En. Compared with a LLaMA-3-8B autoregressive (AR) model trained on the same fine-tuning data, the EV system ties on En→Zh and leads on Zh→En; an oracle-length diagnostic further shows that, in this masked diffusion MT setting, deciding which tokens to reveal first matters less than how the target length is supplied.
Diffusion language models (DLMs) accelerate generation by decoding multiple tokens in parallel, but pushing them to fast decoding often sacrifices accuracy. We attribute this to a static-to-dynamic mismatch: training optimizes local denoising under fixed schedules, while fast inference skips steps and encounters irregular mask patterns. Existing accelerators largely address only one side of this problem. Training-free inference methods speed up decoding without changing the model, leaving the mismatch unresolved; training-side distillation methods adapt the model, but often tie the student to a specific decoding trajectory, limiting composability and still suffering accuracy loss. We propose CD4LM, a composable framework for accurate DLM acceleration. Its core training component, Discrete-Space Consistency Distillation (DSCD) trains a trajectory-invariant student on diverse masked states, reducing the mismatch without fixing a decoding policy. We pair DSCD with a lightweight ConfidenceAdaptive Decoder, while keeping the trained student compatible with external inference-side accelerators. On GSM8K, CD4LM preserves accuracy while achieving a 5.18x speedup; stacking CD4LM with per-step cache accelerators further pushes speedup to 9.73x at teacher-level accuracy, suggesting a scalable path for combining training-side robustness with future decoding advances. Code is available at https://github.com/yihao-liang/CDLM.
Yihao Liang, Ze Wang, Ximeng Sun +7
Princeton University · Advanced Micro Devices, Inc.
One-to-many machine translation (MT) is computationally expensive for autoregressive (AR) systems, which suffer from linear latency scaling with both sequence length and the number of target languages. We explore how diffusion can enable multilingual translation with a discrete diffusion framework that refines all target languages in parallel, achieving sublinear latency scaling with the number of targets, and supports deployment as a single unified model to replace multiple independent systems. Conditioned on a continuous semantic anchor rather than source tokens, our framework supports zero-shot transfer to unseen source languages without retraining, maintaining approximately 75% of its supervised translation quality on zero-shot sources. We investigate the quality-latency frontier and find that with accelerated sampling, it achieves comparable supervised quality to AR baselines with a 2× speedup and 11.9% better zero-shot BLEU. These results highlight the potential of joint multi-stream diffusion as a practical and flexible alternative for efficient one-to-many translation.