cs.CLAug 23, 2026

Length-Adaptive Decoding for Masked Diffusion Machine Translation

Authors: Yan Zhan, Mengkai Hou, Wanting Zhang, Zhijun Gao

Organizations: Peking University · BYD Company Limited

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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→\toZh, Zh→\toEn, and En→\toDe. Our diagnostics show that denoising-friendly lengths need not match reference lengths. Evaluation by three translation experts supports the En↔\leftrightarrowZh adequacy gains, with stronger evidence on Zh→\toEn. Compared with a LLaMA-3-8B autoregressive (AR) model trained on the same fine-tuning data, the EV system ties on En→\toZh and leads on Zh→\toEn; 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.

Figures & tables

Figure 1: Three En →\to Zh cases decoded with 32 denoising steps. A short canvas can drop an action verb or a required placeholder; EV chooses a longer canvas in these two cases and recovers the missing content. The third case shows the opposite boundary: the best compression lies below the EV candidate range, so the ratio baseline is better.
Figure 2: Entropy-Valley method overview. Given a source sentence, EV uses five candidate ratios from the fixed set R\mathcal{R} , yielding up to five distinct target canvases after duplicate lengths are removed. It then probes each candidate length once with an all-mask forward pass through the frozen LoRA-SFT LLaDA-8B model and records the mean predictive entropy Hˉ(L)\bar{H}(L) over all slots except the designated EOS slot. EV selects the entropy valley, L⋆=arg⁡min⁡LHˉ(L)L^{\star}=\arg\min_{L}\bar{H}(L) , and decodes only this selected canvas with minimum-entropy decoding (MED) for T=32T{=}32 steps. The procedure adds no parameters, does not use the reference length of the current sentence, and requires no length predictor.
Figure 3: Length-adaptive decoding for masked-diffusion MT. (A) Mean COMET-22 over three runs for a reference-length upper bound, EV, and a corpus-ratio baseline across three directions. EV closes 64.9%64.9\% on En →\to Zh , 65.3%65.3\% on Zh →\to En , and 33.0%33.0\% on En →\to De . (B) Matched-data AR calibration across three directions. (C) Controlled En →\to Zh diagnostic: with reference target lengths fixed, the tested source-guided unmasking schedules span 0.00810.0081 COMET. With the schedule fixed to MED, three length choices span 0.02650.0265 COMET. The order-schedule sweep uses reference lengths, so it is a diagnostic rather than a deployable alternative to EV.
  COMET-22sacreBLEU
DirectionMethodMean ±\pm StdGap ClosureMean ±\pm StdGap Closure
En →\to ZhOracle length0.8610±0.00130.8610\pm 0.0013100%40.81±0.2340.81\pm 0.23100%
Fixed ratio 0.80.8345±0.00170.8345\pm 0.00170% (baseline)36.72±0.2236.72\pm 0.220%
Entropy-Valley0.8517±0.0006\mathbf{0.8517\pm 0.0006}64.9±7.4%\mathbf{64.9\pm 7.4\%}38.57±0.13\mathbf{38.57\pm 0.13}45.3%
Zh →\to EnOracle length0.8519±0.00070.8519\pm 0.0007100%27.93±0.2327.93\pm 0.23100%
Fixed ratio 1.20.8266±0.00100.8266\pm 0.00100% (baseline)23.65±0.2123.65\pm 0.210%
Table 1: Main LLaDA-8B+LoRA results on WMT22 with 32-step MED decoding. Values are mean ±\pm std over three runs. Length Oracle uses the reference target length at decode time and is an upper bound. Gap closure is (method−Ratio)/(Oracle−Ratio)(\mathrm{method}-\mathrm{Ratio})/(\mathrm{Oracle}-\mathrm{Ratio}) . Paired tests are in Appendix Table 11 .
SystemSizeSettingEn →\to ZhZh →\to EnEn →\to De
LLaDA-8B masked diffusion
Fixed ratio8BLoRA-SFT83.4582.6671.70
Entropy-Valley8BLoRA-SFT85.1784.3172.40
Oracle length †8BLoRA-SFT86.1085.1973.82
Matched-data autoregressive baseline
LLaMA-3-8B + LoRA-SFT8BLoRA-SFT85.07±0.0385.07{\pm}0.0383.68±0.0783.68{\pm}0.0781.19±0.11\mathbf{81.19{\pm}0.11}
Table 2: Matched-data COMET-22 calibration on WMT22. LLaDA rows show the Table 1 means over three runs scaled ×100\times 100 ; the matched AR row reports mean ±\pm std over three runs after training on the same fine-tuning data and evaluating on the same test sets. † Oracle length supplies the reference target length at decode time and is an upper bound, not a deployable method.
Figure 4: Cross-backbone scope check using COMET-22 gap closure (%) for three masked-diffusion backbones and three directions. Gap closure is measured between Ratio and Length Oracle for each backbone–direction pair. Bars show means over three runs; error bars show standard deviations across runs. Full COMET-22 and sacreBLEU values are in Appendix G (Tables 17 and 18 ).
Figure 5: Quality need not track reference length. Target length error and COMET-22 on En →\to Zh WMT22 ( N=2037N{=}2037 ). Points show sentence scores; MAE values use the same outputs, and centroid heights use the main results in Table 1 .
DiagnosticNRatioEV
Placeholder retention5866.9%89.1%
Number retention29077.6%81.3%
r≥0.8r{\geq}0.8 COMET10050.80560.8402
r<0.6r{<}0.6 COMET2360.84490.8503
Table 3: Coverage and length buckets on En →\to Zh. Retention rows measure whether literal source items such as placeholders and numbers are preserved. COMET rows group sentences by rr , the reference target length divided by the source length under the LLaDA tokenizer. Appendix D gives the definitions, confidence intervals, and Zh →\to En companion rows.
DirectionΔ\Delta AdequacyΔ\Delta FluencyPreference
 EV −- RatioEV −- RatioEV / Ratio / Tie
En →\to Zh+0.18+0.18+0.18+0.1828 / 19 / 53
Zh →\to En+0.50\mathbf{+0.50}+0.11+0.1145 / 25 / 30
Table 4: Human evaluation on 100100 stratified WMT22 sentences per direction. Score differences are EV −- Ratio, averaged over three bilingual expert annotators; preference is the per-sentence majority vote.
Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

HyperparameterValue
OptimizerAdamW
AdamW betas(0.9,0.95)(0.9,0.95)
Weight decay0.010.01
Peak learning rate2×10−42{\times}10^{-4}
SchedulerCosine decay
Warm-up5%5\% of training steps
Appendix
Table 5: Training and adaptation hyperparameters for the LLaDA and matched AR LoRA-SFT runs.
RangeKRatiosMAEsBLEUCOMET
En →\to Zh (N=200)
Reported (narrow)5[0.70, 0.90]2.7138.730.8562
Medium7[0.60, 1.00]2.8438.490.8594
Wide9[0.50, 1.10]3.0638.170.8588
Full12[0.50, 1.20]3.2237.580.8577
Extra15[0.40, 1.50]3.8935.980.8476
Appendix
Table 6: Candidate-range control for Entropy-Valley on fixed 200-sentence WMT22 diagnostic subsets. The WMT19 training-corpus median provides a reference scale, while diagnostic range comparisons on WMT22 subsets informed the reported ratio grid. The En →\to Zh grid minimizes length MAE and gives the highest sacreBLEU, while the medium range gives the highest subset COMET; the En →\to De grid gives the highest subset COMET. The Zh →\to En subset benefits from a wider range.
MethodsacreBLEUCOMET-22
En →\to Zh (N=500, 32 steps)
Oracle40.400.8617
Fixed 0.7031.730.8114
Fixed 0.7534.590.8208
Fixed 0.8035.840.8281
Fixed 0.8536.840.8354
Appendix
Table 7: Fixed-ratio sweep on fixed 500-sentence WMT22 diagnostic subsets. EV is compared with fixed ratios around the corpus median for all three directions. On Zh →\to En , the best post-hoc fixed ratio is 1.41.4 at 0.83900.8390 COMET-22; EV reaches 0.84530.8453 , exceeding it by +0.0063+0.0063 COMET and closing 35.4%35.4\% of the oracle-length gap on this subset.
MethodsacreBLEUCOMET-22Cost
Oracle length40.400.86171.00×1.00\times
Fixed ratio 0.835.840.82811.00×1.00\times
EV probe+decode38.480.85571.15×1.15\times
Decode 5 neighbours37.850.8404∼3.5×\sim 3.5\times
Appendix
Table 8: Multi-candidate length selection on the same En →\to Zh WMT22 subset used in Table 7 (N=500, 32 steps). The alternative decodes five candidate canvas lengths obtained as integer offsets around the EV-selected length, {LEV−2,LEV−1,LEV,LEV+1,LEV+2}\{L_{\text{EV}}{-}2,L_{\text{EV}}{-}1,L_{\text{EV}},L_{\text{EV}}{+}1,L_{\text{EV}}{+}2\} (with length collisions deduplicated), to completion and selects by average log-probability; its cost counts all completed decodes.
Dir.MethodCOMETEV −- methodCallsSlotsLatency (s)
En →\to ZhRatio0.8337210.833721+0.010930+0.01093032.0032.00512.66512.660.6440.644
En →\to ZhEV0.844651\mathbf{0.844651}–36.0936.09588.32588.320.7240.724
En →\to ZhDAEDAL0.8280430.828043+0.016608  ([0.012954,0.020210]; p<2×10−4)+0.016608\;([0.012954,0.020210];\ p<2{\times}10^{-4})34.5834.581175.531175.531.7981.798
En →\to ZhCAL0.8378970.837897+0.006754  ([0.003726,0.009880]; p<2×10−4)+0.006754\;([0.003726,0.009880];\ p<2{\times}10^{-4})39.2539.25628.26628.260.7730.773
Zh →\to EnRatio0.8181690.818169+0.016416+0.01641632.0032.00618.43618.430.6350.635
Zh →\to EnEV0.834585\mathbf{0.834585}–36.4336.43708.98708.980.7210.721
Appendix
Table 9: Direct diffusion length comparisons on WMT22 ( N=2037N{=}2037 per direction). Confidence intervals and paired bootstrap pp -values are shown for DAEDAL and CAL. Calls and Slots are per-sentence averages; latency is measured in seconds per sentence in the reported setup. EV is above DAEDAL in both directions and CAL on En →\to Zh. The EV and CAL interval on Zh →\to En includes zero. EV uses fewer calls and slots than CAL, with lower latency in this setup.
Figure 6: Cost–quality Pareto. COMET-22 vs decoding steps T∈{2,4,8,16,32,64,128}T\in\{2,4,8,16,32,64,128\} on WMT22 ( N=2037N{=}2037 , one analyzed run). (A) En →\to Zh: EV is at least as high as Ratio at every TT . (B) En →\to De: EV is higher at small TT and slightly lower than Ratio at T≥64T\geq 64 . Oracle is a length oracle, not a decode oracle.
 En →\to ZhEn →\to De
StepsOracleEVRatioGap%OracleEVRatioGap%
20.69750.70820.6861194%0.39420.40300.3915426%
40.78560.78510.764498%0.50740.51120.4912123%
80.82900.82640.805889%0.60600.60890.5862115%
160.85200.84460.826571%0.68730.67520.668336%
320.85920.85210.832573%0.74170.72430.720916%
Appendix
Table 10: Full cost–quality Pareto on WMT22 ( N=2037N{=}2037 , one analyzed run). Gap% is computed relative to the Ratio and Oracle length columns at the same step budget; values can exceed 100% because Oracle fixes reference length but does not optimize decoding under small TT .
Dir.Comp.Δ\Delta95% CIBoot. ppWilc. pp
En →\to Zh (A)EV–Ratio+0.0196+0.0196[+0.0166,+0.0226][+0.0166,+0.0226]<10−4<10^{-4}4.1×10−324.1{\times}10^{-32}
En →\to Zh (A)Oracle–EV+0.0071+0.0071[+0.0046,+0.0096][+0.0046,+0.0096]<10−4<10^{-4}1.3×10−141.3{\times}10^{-14}
En →\to Zh (A)Oracle–Ratio+0.0266+0.0266[+0.0233,+0.0300][+0.0233,+0.0300]<10−4<10^{-4}3.4×10−583.4{\times}10^{-58}
En →\to De (A)EV–Ratio+0.0034+0.0034[−0.0016,+0.0086][-0.0016,+0.0086]0.0910.0918.3×10−38.3{\times}10^{-3}
En →\to De (A)Oracle–EV+0.0173+0.0173[+0.0121,+0.0225][+0.0121,+0.0225]<10−4<10^{-4}1.6×10−91.6{\times}10^{-9}
En →\to De (A)Oracle–Ratio+0.0208+0.0208[+0.0155,+0.0261][+0.0155,+0.0261]<10−4<10^{-4}6.0×10−146.0{\times}10^{-14}
Appendix
Table 11: Sentence-level paired COMET-22 tests for LLaDA on WMT22 ( N=2037N{=}2037 throughout). Positive Δ\Delta means the first method is better. Paired bootstrap uses 10k resamples; Wilcoxon uses the signed-rank test with Pratt handling of zeros. The table reports run A for En →\to Zh and En →\to De , plus all three Zh →\to En runs. Oracle rows show the run A gap to the reference-length upper bound.
DirectionMethodTProbe capFwd capCOMET-22sacreBLEU
En →\to Zh, one analyzed run, WMT22 N=2037
En →\to ZhRatio 0.8320320.832536.53
En →\to ZhEntropy-Valley32≤5\mathbf{\leq 5}≤37\mathbf{\leq 37}0.8521\mathbf{0.8521}38.56\mathbf{38.56}
En →\to ZhRatio 0.8370370.833436.66
En →\to ZhRatio 0.8400400.833336.71
Zh →\to En, one analyzed run, WMT22 N=2037
Appendix
Table 12: Forward-pass budget control. Ratio is evaluated at the default budget ( T=32T{=}32 ), EV’s maximum total forward pass budget ( T=37T{=}37 ), and a larger budget ( T=40T{=}40 ). Duplicate integer candidate lengths can make EV’s realized budget smaller than the cap.
CategoryNRatioEVΔ\Delta (pp)95% CI (pp)
En →\to Zh — error-type preservation rate
Placeholder (#PRS_ORG#)5866.9%89.1%+22.1 ∗[+12.4, +32.5]
Number / percentage / decimal29077.6%81.3%+3.7 ∗[+1.4, +6.1]
Capitalized named entity (heuristic)87911.2%11.9%+0.7 ∗[+0.0, +1.4]
En →\to Zh — length-compression buckets (LLaDA token ratio)
BucketNRatio COMETEV COMETΔ\Delta COMET95% CI
Appendix
Table 13: Coverage and error categories. Top: source-side literal retention rate, averaged over three runs. Middle: En →\to Zh COMET-22 grouped by r=∣L∗∣/∣x∣r=|L^{*}|/|\mathbf{x}| , the reference target length divided by source length under the LLaDA tokenizer. Bottom: Zh →\to En literal retention. Δ\Delta is in percentage points for retention rows; ∗ marks an interval whose unrounded lower endpoint is above zero.
Likely causeNShareMean Δ\Delta COMET
Overshoot beyond grid12921.6%−0.0278-0.0278
Under-compression12621.1%−0.0206-0.0206
Within-grid wrong pick33956.9%−0.0247-0.0247
Very short source20.3%—
Appendix
Table 14: Failure buckets for the 596 En →\to Zh cases where EV trails Ratio in one analyzed run (WMT22, N=2037N{=}2037 ). Buckets are assigned by mutually exclusive rules.
ScheduleCOMET-22Δ\Delta vs MEDPaired bootstrap pp
MED (baseline)0.8597——
SIG-first0.8603+0.0006+0.00060.620.62
Reverse-SIG0.8523−0.0075-0.0075<10−4<10^{-4}
Hybrid MED →\to SIG, p=0.25p{=}0.250.8582−0.0015-0.00150.210.21
Hybrid MED →\to SIG, p=0.50p{=}0.500.8575−0.0022-0.00220.0110.011
Hybrid MED →\to SIG, p=0.75p{=}0.750.8591−0.0007-0.00070.280.28
Appendix
Table 15: Reveal-order controls on WMT22 En →\to Zh ( N=2037N{=}2037 , T=32T{=}32 ) with reference length and the tokens-per-step schedule fixed. The six source-guided variants span 0.00810.0081 COMET. Strict left-to-right and random order are additional controls and both score below MED. A paired bootstrap value is not reported for Strict L2R.
CategoryNSIG-firstMEDΔ\Delta (SIG −- MED)
Dates320.86520.8665−0.0013-0.0013
Enumeration750.85600.8533+0.0027+0.0027
Named entities2800.86270.8629−0.0002-0.0002
Numbers1230.86790.8724−0.0045-0.0045
Overall4000.8640\mathbf{0.8640}0.8644\mathbf{0.8644}−0.0004\mathbf{-0.0004}
Appendix
Table 16: Source-guided order diagnostic on an En →\to Zh challenge subset ( N=400N{=}400 , one analyzed run, T=32T{=}32 MED vs. SIG-first under matched oracle length). SIG-first reveals high-source-dependence slots first; MED is the confidence-first schedule used in the main experiments.
BackboneDirectionOracle COMETRatio COMETEV COMETΔ\Delta EV −- RatioCOMET Gap closure
LLaDA-8B-Base (paper main results, § 4.2 ; reproduced for cross-backbone comparison)
LLaDA-8B-BaseEn →\to Zh0.8610±0.00130.8610{\pm}0.00130.8345±0.00170.8345{\pm}0.00170.8517±0.0006\mathbf{0.8517{\pm}0.0006}+0.0172+0.017264.9±7.4%\mathbf{64.9{\pm}7.4\%}
LLaDA-8B-BaseZh →\to En0.8519±0.00070.8519{\pm}0.00070.8266±0.00100.8266{\pm}0.00100.8431±0.0004\mathbf{0.8431{\pm}0.0004}+0.0165+0.016565.3±0.8%\mathbf{65.3{\pm}0.8\%}
LLaDA-8B-BaseEn →\to De0.7382±0.00900.7382{\pm}0.00900.7170±0.00900.7170{\pm}0.00900.7240±0.0078\mathbf{0.7240{\pm}0.0078}+0.0070+0.007033.0±8.4%\mathbf{33.0{\pm}8.4\%}
Dream-v0-Base-7B (Qwen2.5-7B continued-pretrained; Qwen2 tokenizer, 152152 k vocab)
Dream-BaseEn →\to Zh0.8414±0.00200.8414{\pm}0.00200.8002±0.00160.8002{\pm}0.00160.8239±0.0004\mathbf{0.8239{\pm}0.0004}+0.0238+0.023857.7±3.1%\mathbf{57.7{\pm}3.1\%}
Appendix
Table 17: Cross-backbone COMET-22 results for Entropy-Valley (WMT22, N=2037N{=}2037 , three runs per aggregate cell, T=32T{=}32 MED). Values are mean ±\pm std across runs. COMET-22 gap closure is (EV−Ratio)/(Oracle−Ratio)(\text{EV}-\text{Ratio})/(\text{Oracle}-\text{Ratio}) per run, then mean ±\pm std. EV is above Ratio in all six Dream-Base and DiffuLLaMA aggregate COMET cells; paired tests for every run give p<10−3p<10^{-3} in all 1818 Dream-Base and DiffuLLaMA comparisons, with positive 95% bootstrap confidence intervals. Absolute scores and closures vary by backbone.
BackboneDirectionOracleRatioEV
LLaDA-8B-BaseEn →\to Zh40.81±0.2340.81{\pm}0.2336.72±0.2236.72{\pm}0.2238.57±0.13\mathbf{38.57{\pm}0.13}
LLaDA-8B-BaseZh →\to En27.93±0.2327.93{\pm}0.2323.65±0.2123.65{\pm}0.2125.28±0.33\mathbf{25.28{\pm}0.33}
LLaDA-8B-BaseEn →\to De22.55±0.7722.55{\pm}0.7720.73±0.6820.73{\pm}0.6821.55±0.85\mathbf{21.55{\pm}0.85}
Dream-BaseEn →\to Zh36.30±0.2236.30{\pm}0.2229.37±0.2029.37{\pm}0.2032.83±0.28\mathbf{32.83{\pm}0.28}
Dream-BaseZh →\to En27.24±0.2827.24{\pm}0.2823.56±0.0423.56{\pm}0.0425.78±0.20\mathbf{25.78{\pm}0.20}
Dream-BaseEn →\to De21.64±0.1621.64{\pm}0.1617.31±0.2517.31{\pm}0.2520.06±0.17\mathbf{20.06{\pm}0.17}
Appendix
Table 18: Companion sacreBLEU view of Table 17 (WMT22, N=2037N{=}2037 , three runs per cell). Values are mean ±\pm std across runs.
Evaluation setPoolRatioEVΔ\Delta COMET (95% CI)
WMT19 held out50k0.74280.74280.75140.7514+0.0086 [0.0034,0.0139]+0.0086\ [0.0034,0.0139]
WMT19 held out200k0.73860.73860.74850.7485+0.0100 [0.0050,0.0152]+0.0100\ [0.0050,0.0152]
FLORES devtest50k0.73980.73980.74810.7481+0.0082 [0.0020,0.0143]+0.0082\ [0.0020,0.0143]
FLORES devtest200k0.73150.73150.74030.7403+0.0088 [0.0024,0.0152]+0.0088\ [0.0024,0.0152]
Appendix
Table 19: De →\to Fr results for two nested training pools. WMT19 held out contains N=1512N{=}1512 sentences, and FLORES uses the full N=1012N{=}1012 De →\to Fr devtest. Both pools use a fixed 500500 update budget and one candidate set for the direction. EV exceeds Ratio in all four settings, with positive paired confidence intervals; paired bootstrap and Wilcoxon tests are below 0.050.05 throughout. This is a robustness check under a fixed update budget, not a comparison of scaling with data size.
Figure 7: Human evaluation on En →\to Zh. (A) Per-annotator adequacy boxplots. (B) Sentence-level Δ\Delta COMET-22 vs 3-expert mean Δ\Delta Adequacy (EV −- Ratio); Spearman ρ=0.637\rho=0.637 ( p<0.001p<0.001 ). The corresponding Zh →\to En correlation is reported in Table 20 .
 ABCAgg.
Panel A: En →\to Zh (N=100)
Δ\Delta Adequacy (EV −- Ratio), mean+0.19+0.19+0.17+0.17+0.19+0.19+0.18+0.18
Δ\Delta Fluency (EV −- Ratio), mean+0.18+0.18+0.20+0.20+0.16+0.16+0.18+0.18
Preference EV / Ratio / Tie28/19/5328/16/5630/27/4328/19/53
Wilcoxon pp (Adequacy)0.038∗0.038^{\ast}0.0700.0700.0650.065—
Wilcoxon pp (Fluency)0.0560.0560.035∗0.035^{\ast}0.1350.135—
Appendix
Table 20: Detailed human-evaluation statistics. The first three columns report annotator-level results. The aggregate preference uses per-sentence majority vote; aggregate Zh →\to En adequacy uses the paired Wilcoxon test on the three-expert mean.
More in this topic

Machine Translation

Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.

See all 267 papers

Explore similar work

CardsList

  1. CARVE: Verified Expansion for Variable-Length Generation in Diffusion Language Models

    Aug 31, 2026Wail Bouhedja, Amr Mohamed, Guokan ShangMasked Diffusion ModelsLLM Inference Acceleration

    1MBZUAI · 2Sorbonne Université · 3Ecole Polytechnique

  2. Masked Diffusion Decoding as xx-Prediction Flow

    Jun 27, 2026Weitian Wang, Lianlei Shan, Shubham Rai +2Masked Diffusion ModelsDiffusion Language Model Decoding

    Robert Bosch GmbH, Germany · University of the Chinese Academy of Sciences, China · Ruhr University Bochum, Germany

  3. DiLaDiff: Distilled Latent-Augmented Diffusion for Language Modeling

    May 22, 2026Jean-Marie Lemercier, Tomas Geffner, Karsten Kreis +3Diffusion Language ModelsDiffusion Language Model Distillation

    NVIDIA