Source-linked AI summary
Efficient One-to-Many Translation with Joint Multi-Stream Diffusion
Yiwen Guan, Jacob Whitehill
TL;DR
One-to-many translation is costly for autoregressive systems, whose latency grows with sequence length and target count. PrismDiff uses joint discrete diffusion guided by a semantic anchor; accelerated sampling matches strong AR supervised quality at 2× speedup, with stronger zero-shot BLEU and about 75% supervised performance on unseen source languages.
Problem
Autoregressive one-to-many translation requires sequential decoding and separate passes for target languages, while parallel non-autoregressive methods can face multimodality and resource costs.
Source: §1 · p. 1Method
PrismDiff jointly refines multiple target-language streams with discrete diffusion conditioned on a shared language-agnostic semantic anchor.
Source: §1 · p. 2Results
At 2× speedup, PrismDiff achieves comparable supervised translation quality to strong AR baselines, maintains approximately 75% of supervised performance on unseen source languages, and outperforms AR counterparts in zero-shot BLEU.
Source: §6 · p. 9Takeaways & Limitations
Parallel multi-stream diffusion with a semantic anchor offers a viable, flexible quality-latency trade-off compared with AR systems.
Source: §6 · pp. 8–9Zero-shot transfer relies on LaBSE alignment; poorly represented or distant low-resource languages may degrade it, and the best encoder may depend on decoder architecture.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
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 $2 \times$ 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.
1 Introduction
The introduction frames one-to-many MT as a latency and deployment challenge, then presents PrismDiff, a joint discrete-diffusion approach guided by a language-agnostic semantic anchor.
1 Introduction
- Autoregressive latency scales linearly with sequence length and target count, while separate one-to-one systems require repeated decoding passes.These costs are especially relevant to resource-constrained one-to-many translation.Source: §1 · p. 1
- Non-autoregressive methods generate tokens in parallel, but conditional independence can cause multimodality, and parallel separate models are memory-intensive and inefficient.The introduction situates diffusion as an alternative based on iterative parallel refinement, while noting that its application to MT remains underexplored.Source: §1 · p. 1
- PrismDiff jointly refines multiple target-language streams from a shared semantic anchor, outperforming independent one-to-one systems through implicit cross-lingual regularization.The anchor also supports transfer to unseen source languages and improves translation accuracy.Source: §1 · p. 2
- PrismDiff forms a competitive supervised quality-latency frontier and a stronger source-side zero-shot frontier than autoregressive baselines.Ablations indicate that effective zero-shot transfer requires a high-quality cross-lingual semantic space; adding diffusion steps alone cannot replace the anchor.Source: §1 · p. 2
2 Related Work
Related work contrasts bilingual non-autoregressive methods and one-to-one diffusion with the paper’s joint one-to-many setting, then describes inference-time sampling as a quality-latency control.
2.1 Non-autoregressive Translation
- Bilingual NAT methods offer speed-quality trade-offs, but are not designed for parallel one-to-many generation or shared semantic representations across target streams.Conditional masked language modeling (Ghazvininejad et al., 2019) and glancing training (Qian et al., 2021) are cited as examples; prior multilingual systems can lack deployment flexibility.Source: §2.1 · p. 2
2.2 Diffusion Language Model
- A single diffusion canvas jointly decodes multiple target languages, unlike prior discrete diffusion work focused on one-to-one translation such as XDLM (Chen et al., 2023).Discrete diffusion operates over categorical tokens and supports parallel reconstruction with variable sampling budgets.Source: §2.1 · p. 2, §2.2 · p. 2
2.3 Inference-time Acceleration for Diffusion Models
- A logarithm-uniform inference schedule varies sampling steps to move one trained model across quality-latency points without changing model capacity or retraining.The paper adopts the schedule applied in Irwin et al. 2025; diffusion quality depends on denoising-step count, creating a latency-quality trade-off.Source: §2.3 · p. 3
3 Methodology
The method uses discrete masked diffusion to generate multiple translations from a continuous semantic anchor, processing target streams together while isolating their attention.
3.1 Preliminaries
- The model represents a source sentence and K target translations as a prompt and response, with the prompt unmasked and target sequences progressively corrupted and reconstructed.The response comprises discrete tokens from a vocabulary.Source: §3.1 · p. 3
- Discrete masked diffusion corrupts target responses over timesteps and trains a neural network to predict original tokens at masked positions.Inference starts from a fully masked sequence and iteratively denoises it, remasking a fraction of tokens at each refinement step.Source: §3.1 · p. 3
3.2 Proposed Method
- PrismDiff extracts source hidden states with LaBSE and projects them into the diffusion model’s dimension as a continuous semantic anchor.The encoder is pretrained to align parallel sentences across languages in a shared embedding space.Source: §3.2 · p. 3
- A unified canvas concatenates the anchor with all noisy target streams, which the network processes together to predict masked tokens across streams.Each stream uses a language-specific mask token, and Figure 1 illustrates concurrent generation of multiple targets from one source sentence.Source: §3.2 · pp. 3–4, Figure 1 · p. 4
- Exclusive within-stream attention avoids interference between target languages while shared parameters and joint optimization provide implicit cross-lingual regularization.Reset positional encoding restarts positions for each target stream, allowing target languages to be reordered, removed, or added without affecting performance.Source: §3.2 · p. 4
- At inference, the canvas begins fully masked and retains the most confident tokens at each step, remasking the rest according to a sampling schedule.The method tests vanilla and logarithm-uniform schedules; final outputs also undergo CTC collapse to remove consecutive duplicates and blanks.Source: §3.2 · pp. 4–5
4 Experimental Setup
The experiments compare PrismDiff with autoregressive, independent-diffusion, source-token diffusion, and tree-based baselines across supervised and source-side zero-shot translation. They use Multi30K and Europarl, reporting translation quality and end-to-end latency under controlled model and decoding settings.
4.1 Baselines
- Baselines include decoder-only AR models, independently trained diffusion decoders, source-token-conditioned multi-stream diffusion, and TET, a unified encoder-tree NAT model.Diff-Indep isolates joint multi-stream optimization; Diff-SrcTokens tests anchor conditioning. TET lacks zero-shot transfer. The study does not directly compare one-to-one NAT systems such as GLAT and XDLM.Source: §4.1 · p. 5
4.2 Datasets and Evaluation Metrics
- Multi30K supplies 29k En-{De, Fr} training examples and 1k test examples; Europarl adds 260k training and 33k test examples across four targets.The Multi30K Es-X zero-shot test set is created by translating English into Spanish with NLLB-200; Europarl uses an original human-authored Spanish evaluation set.Source: §4.2 · p. 5
- Quality and semantic transfer are measured with SacreBLEU and COMET, while efficiency is average end-to-end latency in milliseconds for all requested targets on one model instance.The comparison focuses on single-model one-to-many latency; AR replicas can reduce wall-clock latency but increase memory and deployment complexity.Source: §4.2 · p. 5
4.3 Implementation Details
- Diffusion systems use an 8-layer, roughly 230M-parameter Transformer backbone, with T=100 on Multi30K and T=40 on Europarl unless noted.Accelerated sampling varies inference steps; models use shared target-language BPE, CTC collapse after diffusion decoding, and truncation at EOS. LaBSE is frozen while its linear projection adapter is trained.Source: §4.3 · pp. 5–6
5 Results and Analysis
PrismDiff offers a competitive quality–latency frontier against AR systems, with stronger source-side zero-shot performance and benefits from its semantic anchor and joint stream optimization. Encoder choice and deployment settings also affect transfer quality and scaling behavior.
5 Results and Analysis
- PrismDiff reaches comparable SacreBLEU and COMET scores with substantially lower latency than strong AR baselines, while offering advantages in source-side zero-shot transfer and parallel generation.The reported quality–latency comparison analyzes frontiers induced by diffusion sampling steps and AR decoder depth on Multi30K and Europarl.Source: §5 · p. 6, Table 1 · p. 6, Table 2 · p. 7
5.1 Supervised Quality-Latency Frontier
- At roughly 70–100 ms, PrismDiff reaches most supervised BLEU gains within 25–30 sampling steps and remains competitive with AR across a wide latency range.PrismDiff maintains near-peak translation quality (∼37 BLEU) at strict low-latency budgets, while the AR frontier degrades as decoder depth is reduced.Source: §5.1 · p. 6
5.2 Source-Side Zero-Shot Transfer
- On unseen Spanish sources, PrismDiff’s zero-shot frontier consistently outperforms AR baselines, though it does not close the supervised-to-zero-shot gap.Figure 2 distinguishes supervised En-X results from zero-shot Es-X results; its left panel compares AR and PrismDiff, while its right panel compares diffusion systems. The authors speculate that iterative refinement mitigates error accumulation.Source: §5.2 · p. 7, Figure 2 · p. 7
5.3 Ablation Studies
- At T=100, Diff-SrcTokens peaks at 35.07 BLEU versus 33.84 at T=10, but PrismDiff outperforms it at matched latency and preserves strong Es-X performance.Diff-SrcTokens falls below 1 BLEU on Es-X; the authors conclude that increasing T improves the sampling frontier but cannot replace the semantic anchor’s cross-lingual mapping.Source: §5.3 · p. 7
- Exclusive attention performs better than full attention for PrismDiff, while Diff-SrcTokens shows the reverse; standard positional encoding yields 23.8 versus 33.5 BLEU at t=10.The authors suggest cross-stream attention supports implicit alignment without the anchor but may become redundant and introduce interference when the anchor supplies that alignment.Source: §5.3 · pp. 7–8
- At t=30, joint PrismDiff exceeds Diff-Indep by ∼1.5 supervised BLEU and ∼1.4 zero-shot BLEU, suggesting implicit cross-lingual regularization beyond the shared anchor.Diff-Indep uses the same diffusion mechanism and anchor at T=100 but generates each target language independently.Source: §5.3 · p. 8
5.4 Encoder Choice and Anchor Quality
- Within PrismDiff, LaBSE, mBERT, and XLM-R perform comparably on supervised En-X, while zero-shot Es-X performance diverges: LaBSE nears 28 BLEU, XLM-R 13, and mBERT 10, respectively.Under traditional AR decoding, mBERT reaches ∼39 supervised BLEU but ∼8.5 BLEU on Es-X zero-shot transfer.Source: §5.4 · p. 8
5.5 Deployment Flexibility and Scaling
- PrismDiff is invariant to target generation order, loses less than 1% quality when target streams are omitted at inference, and exhibits sublinear latency growth as target count increases.The authors contrast this scaling with AR systems’ strict linear latency growth when generating targets sequentially.Source: §5.5 · p. 8
6 Conclusion
PrismDiff uses a language-agnostic semantic anchor for parallel one-to-many translation, with controllable quality–latency trade-offs and source-side zero-shot transfer. Accelerated sampling gives comparable supervised quality to strong AR baselines with a 2× speedup, while zero-shot performance reaches approximately 75% of supervised performance.
6 Conclusion
- PrismDiff is a discrete diffusion framework for parallel one-to-many translation, guided by a language-agnostic semantic anchor.The paper presents it as a way to explore the design space of parallel multilingual machine translation.Source: §6 · pp. 8–9
- With accelerated sampling, PrismDiff achieves comparable supervised translation quality to strong AR baselines with a 2× speedup and maintains approximately 75% of supervised performance on unseen source languages.It also outperforms AR counterparts in zero-shot BLEU.Source: §6 · p. 9
- The authors identify possible extensions to other one-to-many or many-to-many generation tasks and more robust anchor alignment strategies.These are proposed future directions, not evaluated results.Source: §6 · p. 9
Limitations
PrismDiff has a controllable quality–latency frontier, but does not uniformly surpass strong AR models in supervised quality and remains slower than single-step NAR decoding. Its results also depend on the pretrained encoder and implementation choices, while its efficiency evaluation is limited to single-GPU, batchsize=1 latency.
Limitations
- PrismDiff does not uniformly outperform strong AR baselines on supervised translation quality, and diffusion decoding remains slower than single-step NAR models.Fewer refinement steps reduce latency, while additional steps enhance quality.Source: Limitations · p. 9
- Zero-shot transfer may degrade for poorly represented or truly distant low-resource languages in the pretrained encoder, and encoder choice can depend on the decoder.The authors limit their conclusion about LaBSE to PrismDiff rather than all translation decoders.Source: Limitations · p. 9
- Fixed language block lengths of 40 for Multi30K and 80 for Europarl can overflow on longer sentences, contributing considerably to poor Europarl performance.The authors suggest that tuning implementation strategies and hyperparameters may improve performance.Source: Limitations · p. 9
- Adding cross-stream mechanisms such as gated cross-attention could introduce more computation and inter-stream interference.The paper presents these mechanisms as a possible direction, not an evaluated improvement.Source: Limitations · p. 9
- Efficiency evaluation uses single-GPU, single-instance, batchsize=1 latency; optimizing batch throughput remains future work.The authors associate this setting with serving scenarios where synchronous batching is impractical.Source: Limitations · p. 9
A.1 Logarithm-Uniform Masking Schedule
The logarithm-uniform masking schedule is discretized over training steps to cover the range of masking ratios. The authors report that this schedule consistently outperforms other simple strategies in their task.
A.1 Logarithm-Uniform Masking Schedule
- The schedule covers the full range of masking ratios with the desired log-uniform distribution.The masking ratio m is bounded by [1/L, 1], and tokens are remasked according to predicted confidence.Source: §A.1 · p. 11
- For fixed-step training, the continuous prior is discretized over the total training steps and implemented deterministically by mapping steps to masking ratios.The example uses L = 40 with K = 4 discrete levels.Source: §A.1 · p. 11
- The logarithm-uniform masking schedule consistently outperforms other simple strategies in this task.Source: §A.1 · p. 11
A.2 Scaling Analysis: Latency and Memory
PrismDiff’s latency grows sublinearly with the number of target streams, while its memory remains manageable across the reported multi-stream settings. The comparison with autoregressive decoding highlights a trade-off between latency and memory scaling.
A.2 Scaling Analysis: Latency and Memory
- PrismDiff costs O(t · K · L^2) over t refinement steps, versus O(K · L^3) for sequential AR decoding, favorable when t ≪ L.Running K AR model replicas in parallel could reduce wall-clock latency but multiplies memory by K; PrismDiff uses one model instance.Source: §A.2 · pp. 11–12
- PrismDiff’s latency scales sublinearly with target streams K; generating K languages costs K× a single-language diffusion pass but K× less than K separate passes.Figure 4 compares PrismDiff with decoder-only AR models, including accelerated sampling at t = 50 or t = 25.Source: §A.2 · p. 11, Figure 4 · p. 12
- With block length l = 40, memory rises less than 10% from 2.66 GB at K = 1 to 2.91 GB at K = 8; growth is more pronounced for longer blocks.The canvas size grows as Lsrc + K × Ltgt, but even K = 8 with l = 300 fits standard GPUs without complex multi-GPU model parallelism.Source: §A.2 · p. 12, Table 3 · p. 12
A.3 Full Results
The appendix details Multi30K evaluations, explains how semantic alignment shapes attention, and illustrates simultaneous refinement and zero-shot behavior. It also describes TET’s implementation and the licenses of the datasets and models used.
A.3 Full Results
- The Multi30K full results cover implementation and architecture setups, sampling budgets, AR decoder-layer counts, and multilingual encoders used to extract the source-side anchor.Tables 5–8 also identify Spanish as a zero-shot source in the architecture results and describe the sampling and encoder comparisons.Source: §A.3 · p. 12, Table 5 · p. 14, Table 6 · p. 14
A.4 The Role of Cross-Lingual Alignment in Attention
- With a semantic anchor, exclusive attention lets each target language denoise independently; without one, full cross-stream attention can supply auxiliary context for source-token-conditioned models.The passage attributes the difference to the anchor’s cross-lingual alignment and the surface-form interference that direct target-stream attention can introduce.Source: §A.4 · p. 12
A.5 Inference Examples
- PrismDiff refines all target languages simultaneously at each inference step; the displayed t = 0 prediction is CTC-collapsed to remove repetitive tokens when present.The appendix includes accelerated Multi30K and Europarl translation-process examples.Source: §A.5 · pp. 12–13, Table 9 · p. 16, Table 10 · p. 16
A.6 Zero-Shot Transfer Examples
- In qualitative Spanish-to-French zero-shot examples, AR baselines show error cascading, hallucination, structural collapse, and non-existing words.The models were trained on English-to-German and English-to-French; the examples use Spanish as input and show French outputs.Source: §A.6 · p. 13, Table 11 · p. 17
A.7 Implementation of TET
- TET generates all target languages in one CTC-decoding pass, and its higher parameter count does not proportionally increase latency because of its parallel tree structure.The Europarl TET model has four target-language leaf nodes, depth six, 16 Transformer encoder layers, and 556M parameters.Source: §A.7 · p. 13
A.8 Licenses and Usage of Artifacts
- The appendix reports public dataset licenses and model licenses, including NLLB-200’s CC-BY-NC 4.0 license, which limits use to non-commercial research.It states that the datasets and models are used consistently with their intended research purposes.Source: §A.8 · p. 13