Pith. sign in

REVIEW 4 major objections 5 minor 42 references

ViPo-MLLM: Visual-Pose Multimodal LLM for Gloss-Free Sign Language Translation

T0 review · 4 major / 5 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fusing RGB video with body-and-hand pose through cross-modal attention lets a language model translate sign language without gloss labels and reach new gloss-free state-of-the-art scores.

desk verdict Solid engineering SOTA on two gloss-free SLT benchmarks via RGB+pose fusion into an LLM; the gains are real but the causal credit for pose/CMA is thinner than the abstract claims. read the letter →

arxiv 2607.03657 v1 pith:5D45XYU5 submitted 2026-07-04 cs.CV cs.AI

Ahmed Abul Hasanaath , Bicheng Xu , Mir Rayat Imtiaz Hossain , Leonid Sigal , Hamzah Luqman This is my paper · ORCID
classification cs.CVcs.AI
keywords SignLanguageTranslationGloss-freeSLTMultimodalLLMsPosefeaturesCross-modalattentionVisual-languagealignmentPHOENIX14TCSL-Daily
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Gloss-free sign language translation aims to turn continuous signing videos into fluent spoken sentences without expensive intermediate word-like gloss annotations. The authors argue that existing single-stream or weakly fused visual features miss the fine articulatory detail of hands, body, and motion that carries linguistic meaning. They introduce ViPo-MLLM, which extracts complementary spatio-temporal RGB features and intermediate pose features, models each stream’s local dynamics, then fuses them with time-aware cross-modal attention. The fused sequence is wrapped in a structured prompt and decoded by a fine-tuned large language model trained with both contrastive alignment and ordinary language-modeling loss. On the two standard benchmarks the method sets new gloss-free records and remains competitive with several gloss-supervised systems, showing that carefully fused pose and appearance cues can close much of the gap that glosses once filled.

What carries the argument

The multimodal fusion block: Intra-Modal Temporal Modeling (shared TCNs) followed by Cross-Modal Temporal Modeling (a 4-layer Cross-Modal Attention module) that lets pose and RGB streams exchange long-range information before the fused tokens enter the language model.

What would settle it

On the same PHOENIX14T and CSL-Daily test sets, replace the intermediate OpenPose features with pure RGB or with final heatmap keypoints while keeping every other component fixed; if BLEU-4 and ROUGE-L fall back to or below the previous gloss-free best numbers, the claimed necessity of the pose-plus-cross-modal design is refuted.

Watch

Extended reading notes

Core claim

When spatio-temporal RGB features and intermediate OpenPose body-and-hand features are first processed by shared temporal convolutions and then fused by a four-layer cross-modal attention module, the resulting representation, conditioned by a short instructive prompt and contrastively aligned with text, enables a large language model to produce gloss-free translations that surpass previous gloss-free systems and approach gloss-based performance on both German and Chinese continuous signing benchmarks.

Load-bearing premise

That intermediate pose features plus a frozen RGB encoder already contain enough articulatory information for fluent translation, even without explicit facial-expression modeling or any gloss supervision.

Editorial extensions

If this is right

  • Gloss annotation pipelines can be bypassed for continuous SLT while still matching or exceeding many gloss-supervised baselines.
  • Pose streams extracted once by off-the-shelf estimators become a reusable, compact linguistic channel that can be fused with any strong visual encoder.
  • Structured instructive prompts plus contrastive video-text alignment form a lightweight recipe for adapting large language models to continuous visual languages.
  • The same fusion pattern can be applied to other multi-articulatory visual languages where hand shape, body posture, and motion must be jointly decoded.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because facial cues are never modeled explicitly, performance on languages that rely heavily on non-manual markers may still lag until a third stream is added.
  • The two-stage training schedule (contrastive warm-up then joint loss) suggests that pure language-modeling gradients alone are insufficient to align continuous visual tokens with text.
  • If intermediate pose features already carry most of the articulatory signal, lighter pose estimators or even synthetic pose could further reduce inference cost for real-time deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents ViPo-MLLM, a gloss-free sign language translation system that extracts spatio-temporal RGB features with a frozen USTM encoder and pose features via intermediate OpenPose representations, fuses them through shared-TCN Intra-Modal Temporal Modeling and a multi-layer Cross-Modal Attention (CMA) block, then conditions an mT0-XL LLM (LoRA-tuned) with a structured instructive prompt plus few-shot examples. Training combines a contrastive alignment loss (pooled multimodal features vs. English translations of ground-truth sentences) with standard cross-entropy language modeling. On PHOENIX14T and CSL-Daily the method reports new gloss-free SOTA BLEU-4 / ROUGE-L scores (27.10 / 51.50 and 25.85 / 54.53) that are competitive with several gloss-based systems; ablations examine RGB vs. pose, CMA depth, LLM backbone, and prompt style.

Significance. If the performance gains are robustly attributable to the proposed fusion, the work would strengthen the case that gloss-free SLT can approach gloss-based accuracy by combining a strong frozen visual backbone with compact pose cues and LLM decoding, reducing reliance on expensive gloss annotation. Public code, two standard benchmarks, and systematic component ablations (Tables 2–5) are concrete strengths that aid reproducibility and allow the community to test the claims. The result is incremental rather than paradigm-shifting, but useful for the multimodal SLT literature.

major comments (4)
  1. Abstract and §4.2 claim that the pose cues and cross-modal attention are the key drivers of the new gloss-free SOTA (Table 1: PHOENIX14T BLEU-4 27.10, CSL-Daily 25.85). Table 3 shows that CMA depth 0 already reaches BLEU-4 25.67 (near SpaMo/MMSLT), while Table 2 shows RGB alone at 25.54 and RGB+pose without contrastive at 26.61. The incremental gains from the proposed IMTM/CMTM stack are therefore modest; without a controlled experiment that freezes the identical USTM backbone and LLM and varies only the fusion module (or replaces OpenPose with a null pose stream), the causal contribution of pose+CMA remains under-supported relative to the headline claim.
  2. §1 and the Abstract repeatedly state that fine-grained modeling of “hands, body, and facial cues” is required and that ViPo-MLLM supplies it. Section 3.2 describes only full-body and hand pose features (OpenPose intermediate maps, spatially pooled) plus RGB; no face landmarks, mouthing stream, or facial-expression head is used. Because non-manual facial signals are linguistically critical in both German and Chinese Sign Language, the claim that the extracted features constitute complete articulatory information is incomplete and should be either substantiated (e.g., by an explicit face ablation) or qualified.
  3. All reported numbers (Tables 1–5) appear to be single-run point estimates with no standard deviations, multiple random seeds, or statistical significance tests. Given that the absolute BLEU-4 margins over the strongest gloss-free baselines are 1–4 points, the absence of uncertainty quantification makes it impossible to judge whether the SOTA ranking is stable; at minimum the main Table 1 entries should be accompanied by mean±std over ≥3 seeds.
  4. §3.3 defines the contrastive target as “English translations of the ground-truth sentences” even though the target languages are German (PHOENIX14T) and Chinese (CSL-Daily). While the multilingual LLM can in principle handle this, the design choice is never justified or ablated against using the original-language references; it risks injecting an unnecessary translation bottleneck into the alignment objective and should be clarified or replaced.
minor comments (5)
  1. Title and running header contain spacing artifacts (“TRANSLA TION”, “LITERA TURE”); clean for camera-ready.
  2. Figure 1 is dense; the distinction between frozen USTM stages, discarded MS-TCN/BiLSTM heads, and the trainable IMTM/CMTM path is hard to parse at a glance—consider a clearer legend or simplified schematic.
  3. §3.2 states that the same TCN is shared across modalities “to enforce consistent temporal patterns”; this design choice is reasonable but never ablated against modality-specific TCNs, so the claim remains an untested axiom.
  4. Table 1 places ViPo-MLLM under the “Gloss-free” block while the text (§4.2) correctly notes that the RGB encoder is pretrained; a short footnote clarifying the weakly-supervised status of the visual backbone would avoid reader confusion.
  5. Prompt ablation (Table 5) uses only three few-shot examples whose selection criteria are not described; a brief note on how the examples were chosen would improve reproducibility.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: empirical SOTA on held-out public benchmarks; only a minor non-load-bearing self-citation of the frozen USTM RGB encoder.

  1. self citation load bearing [Section 3.2 Sign Feature Extraction / Spatio-Temporal Features]
    "We extract spatio-temporal features using the USTM framework, which employs a hierarchical vision encoder with stacked spatio-temporal modules combining Swin Transformer stages and lightweight temporal adapters. We use pretrained recognition weights and keep the encoder frozen, discarding late-stage components (MS-TCN, BiLSTM, classification heads)..."

    USTM is the authors’ own concurrent recognition model (arXiv:2512.13415). Using its frozen pretrained weights injects a self-cited component into the pipeline. The citation is not load-bearing for the SOTA claim, however: the SLT metrics remain external test-set evaluations, and the paper’s novel fusion/prompting contributions are still independently ablated.

full rationale

ViPo-MLLM is a standard empirical multimodal systems paper. Its central claims are measured BLEU/ROUGE numbers on the official held-out test splits of PHOENIX14T and CSL-Daily (Table 1). Training optimizes a contrastive loss against English translations of the ground-truth sentences plus ordinary cross-entropy language-modeling loss on the target spoken sentences; neither loss, nor any free parameter, is fitted to the test metrics that are later reported as “predictions.” The architectural choices (IMTM TCN, 4-layer CMA, structured prompt, LoRA on mT0-XL) are described and ablated independently (Tables 2–5). The sole self-citation of note is the frozen USTM RGB encoder taken from the authors’ concurrent recognition work; this supplies a strong pretrained visual backbone but does not force the reported SLT numbers by construction, uniqueness import, or definitional identity. No self-definitional loop, no fitted-input-called-prediction, no uniqueness theorem, and no renaming of a known result appear. The derivation chain is therefore self-contained against external benchmarks, warranting only a minimal score for the modular self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central performance claim rests on standard deep-learning practice plus a handful of architectural and training choices (CMA depth 4, LoRA rank 16, two-stage schedule, instructive prompt, frozen USTM/OpenPose). No new physical constants or free parameters are fitted to the test metrics themselves; the free parameters are ordinary hyperparameters selected on validation performance.

free parameters (5)
  • CMA depth = 4
    Chosen by ablation (Table 3); peak at depth 4. Directly affects the reported SOTA numbers.
  • LoRA rank = 16
    Set to r=16 for mT0-XL fine-tuning; not ablated.
  • contrastive temperature τ = learnable
    Learnable temperature in the InfoNCE-style loss; value not reported.
  • TCN kernel size / layers = kernel=5, 2 layers
    2-layer 1D TCN with kernel size 5 and max-pooling; fixed design choice.
  • training schedule (warm-up + joint epochs) = 100+50
    100 contrastive-only epochs then 50 joint epochs with Reduce-on-Plateau; chosen by authors.
assumptions (4)
  • domain assumption Frozen pretrained USTM RGB encoder and OpenPose intermediate features supply sufficient visual and articulatory information for gloss-free translation.
    Section 3.2; encoder kept frozen, late stages discarded.
  • domain assumption Contrastive alignment of temporally-pooled multimodal features to frozen LLM embeddings of English translations improves subsequent generation.
    Learning Objective subsection; two-stage training.
  • domain assumption BLEU-n and ROUGE-L on PHOENIX14T / CSL-Daily are adequate proxies for translation quality.
    Standard SLT evaluation practice used throughout Section 4.
  • ad hoc to paper Shared TCN weights across RGB and pose enforce consistent temporal patterns that aid fusion.
    IMTM description in Multimodal Feature Fusion.
invented entities (1)
  • ViPo-MLLM fusion stack (IMTM + CMTM/CMA)
    purpose: Explicitly model intra-modal dynamics and cross-modal long-range dependencies between RGB and pose streams before LLM decoding.
    Architectural construct introduced in Section 3; no independent physical existence outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViPo-MLLM: Visual-Pose Multimodal LLM for Gloss-Free Sign Language Translation." pith.science (2026). https://pith.science/paper/5D45XYU5

@misc{pith2026260703657,
  author       = {Pith},
  title        = {Pith review of: ViPo-MLLM: Visual-Pose Multimodal LLM for Gloss-Free Sign Language Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5D45XYU5}},
  note         = {Machine review of arXiv:2607.03657}
}
read the original abstract

Gloss-free Sign Language Translation (SLT) translates sign language videos into spoken-language sentences without gloss annotations, avoiding costly labeling but requiring fine-grained modeling of hands, body, and facial cues. Existing methods often use single-modality or weakly fused features, limiting performance. We propose ViPo-MLLM, a framework that integrates spatio-temporal RGB and human pose features. Dedicated encoders model intra-modal dynamics and cross-modal attention captures long-range dependencies. The fused representation is conditioned with a structured prompt and processed by an LLM trained with contrastive and language modeling objectives. The proposed model was evaluated on the PHOENIX14T and CSL-Daily datasets and achieved new state-of-the-art results on both datasets. Moreover, the ViPo-MLLM model attained competitive performance compared to gloss-based recognition approaches, confirming the effectiveness of the proposed pose cues and cross-modal attention mechanisms.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

  1. [1]

    Sign Language Trans- lation (SLT) aims to bridge communication gaps by trans- lating sign-language videos into spoken sentences

    INTRODUCTION Sign language conveys meaning through hand gestures, body movements, and facial expressions. Sign Language Trans- lation (SLT) aims to bridge communication gaps by trans- lating sign-language videos into spoken sentences. Despite recent progress, SLT remains challenging due to subtle tem- poral dynamics and the modality gap between visual inp...

  2. [2]

    Prior work ad- dresses this throughtemporal and semantic modelingand cross-modal alignment

    LITERA TURE REVIEW Gloss-Free SLT.Gloss-free SLT translates sign language videos directly into spoken sentences without glosses, but often underperforms gloss-based methods. Prior work ad- dresses this throughtemporal and semantic modelingand cross-modal alignment. TSPNet [8] models hierarchical temporal semantics, while CSGCR [9] uses cross-modal re- ran...

    arXiv 2026

  3. [3]

    SignLLM [5] maps sign videos to discrete tokens aligned with LLMs

    and FLa-LLM [2] use pretrained models with adapters or factorized training to improve visual representations without gloss supervision. SignLLM [5] maps sign videos to discrete tokens aligned with LLMs. SpaMo [13] emphasizes spatial and motion cues, and MMSLT [4] exploits multimodal LLMs to generate detailed textual descriptions aligned with video. Multim...

  4. [4]

    Re- cent work leverageslarge-scale pretraining and multimodal LLMs

    jointly models spatial and temporal cues.Structured interactionuses graph-based encodings (MSeqGraph [19]) and skeleton-aware scaling (SANet [20]), while multi-stream encoders like TwoStream-SLT [21] reduce redundancy. Re- cent work leverageslarge-scale pretraining and multimodal LLMs. MMSLT [4] aligns video features with textual de- scriptions

  5. [5]

    Translate the given sentence into<language>

    METHODOLOGY 3.1. Framework Overview Figure 1 shows the ViPo-MLLM architecture. Spatio-temporal and pose-based features are extracted from a sign video using a USTM encoder [22] and OpenPose [23], respectively. These features are fused via Intra-Modal and Cross-Modal Tempo- ral blocks to capture local and global dependencies. The fused representation, comb...

  6. [6]

    Experimental Setup Datasets.We evaluate on two benchmark SLT datasets: PHOENIX14T [25] and CSL-Daily [26]

    EXPERIMENTS 4.1. Experimental Setup Datasets.We evaluate on two benchmark SLT datasets: PHOENIX14T [25] and CSL-Daily [26]. PHOENIX14T contains 8,247 German sign language videos with 1,085 glosses (avg. 7.7 per sentence), split into 7,096 training, 519 validation, and 642 test samples. CSL-Daily has 20,654 Chi- nese sign language videos covering 2,000 glo...

  7. [7]

    CONCLUSION We proposed ViPo-MLLM, an SLT framework that fuses spatio-temporal RGB features with pose representations via cross-modal temporal attention. By modeling intra-modal dynamics and cross-modal dependencies, and leveraging structured prompting with contrastive alignment, our method captures fine-grained motion and long-range temporal pat- terns cr...

  8. [8]

    ACKNOWLEDGEMENTS The authors would like to acknowledge the support provided by King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project number ISP24226

Show all 42 references
  1. [9]

    Sign language transformers: Joint end-to-end sign language recognition and translation,

    N. C. Camgoz, O. Koller, S. Hadfield, and R. Bow- den, “Sign language transformers: Joint end-to-end sign language recognition and translation,” inIEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10023–10033

    2020

  2. [10]

    Factorized learning assisted with large language model for gloss-free sign language translation,

    Z. Chen, B. Zhou, J. Li, and et al, “Factorized learning assisted with large language model for gloss-free sign language translation,”arXiv preprint arXiv:2403.12556, 2024

    2024 arXiv

  3. [11]

    Gloss-free sign language translation: Improving from visual-language pretraining,

    B. Zhou, Z. Chen, A. Clap ´es, et al., “Gloss-free sign language translation: Improving from visual-language pretraining,” inIEEE/CVF International Conference on Computer Vision, 2023, pp. 20871–20881

    2023

  4. [12]

    Leveraging the power of mllms for gloss-free sign language transla- tion,

    J. Kim, H. Jeon, J. Bae, and H. Y . Kim, “Leveraging the power of mllms for gloss-free sign language transla- tion,” inIEEE/CVF International Conference on Com- puter Vision, 2025, pp. 21048–21058

    2025

  5. [13]

    Llms are good sign language translators,

    J. Gong, L. G. Foo, Y . He, and et al, “Llms are good sign language translators,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18362–18372

    2024

  6. [14]

    Lost in translation, found in context: Sign language translation with contextual cues,

    Y . Jang, H. Raajesh, L. Momeni, G. Varol, and Andrew Zisserman, “Lost in translation, found in context: Sign language translation with contextual cues,” inComputer Vision and Pattern Recognition Conference, 2025, pp. 8742–8752

    2025

  7. [15]

    Sign2gpt: Leveraging large language models for gloss-free sign language translation,

    R. Wong, N. C. Camgoz, and R. Bowden, “Sign2gpt: Leveraging large language models for gloss-free sign language translation,”arXiv preprint arXiv:2405.04164, 2024

    2024 arXiv

  8. [16]

    Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation,

    D. Li, C. Xu, X. Yu, and et al, “Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation,”Advances in Neural Information Processing Systems, vol. 33, pp. 12034–12045, 2020

    2020

  9. [17]

    Conditional sen- tence generation and cross-modal reranking for sign lan- guage translation,

    J. Zhao, W. Qi, W. Zhou, and et al, “Conditional sen- tence generation and cross-modal reranking for sign lan- guage translation,”IEEE Transactions on Multimedia, vol. 24, pp. 2662–2672, 2021

    2021

  10. [18]

    A token-level contrastive framework for sign language translation,

    B. Fu, P. Ye, L. Zhang, and et al, “A token-level contrastive framework for sign language translation,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  11. [19]

    Gloss attention for gloss-free sign language translation,

    A. Yin, T. Zhong, L. Tang, and et al, “Gloss attention for gloss-free sign language translation,” inIEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2551–2562

    2023

  12. [20]

    Visual alignment pre- training for sign language translation,

    P. Jiao, Y . Min, and X. Chen, “Visual alignment pre- training for sign language translation,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 349–367

    2024

  13. [21]

    An ef- ficient sign language translation using spatial configu- ration and motion dynamics with llms,

    E. J. Hwang, S. Cho, J. Lee, and J. C. Park, “An ef- ficient sign language translation using spatial configu- ration and motion dynamics with llms,”arXiv preprint arXiv:2408.10593, 2024

    2024 arXiv

  14. [22]

    Better sign language translation with stmc-transformer,

    Y . Kayo and J. Read, “Better sign language translation with stmc-transformer,”arXiv preprint, 2020

    2020

  15. [23]

    Lost in translation, found in embeddings: Sign language translation and alignment,

    Y . Jang, L. Momeni, Z. Jiang, et al, and Andrew Zisser- man, “Lost in translation, found in embeddings: Sign language translation and alignment,”arXiv preprint arXiv:2512.08040, 2025

    2025

  16. [24]

    Multimodal sign language recognition via temporal deformable convolu- tional sequence learning.,

    K. Papadimitriou and G. Potamianos, “Multimodal sign language recognition via temporal deformable convolu- tional sequence learning.,” inInterspeech, 2020, pp. 2752–2756

    2020

  17. [25]

    Multi-channel transformers for multi-articulatory sign language translation,

    N. C. Camgoz, O. Koller, S. Hadfield, and R. Bowden, “Multi-channel transformers for multi-articulatory sign language translation,” inEuropean Conference on Com- puter Vision. Springer, 2020, pp. 301–319

    2020

  18. [26]

    Spatial- temporal multi-cue network for sign language recogni- tion and translation,

    H. Zhou, W. Zhou, Y . Zhou, and H. Li, “Spatial- temporal multi-cue network for sign language recogni- tion and translation,”IEEE Transactions on Multimedia, vol. 24, pp. 768–779, 2021

    2021

  19. [27]

    Graph- based multimodal sequential embedding for sign lan- guage translation,

    S. Tang, D. Guo, R. Hong, and M. Wang, “Graph- based multimodal sequential embedding for sign lan- guage translation,”IEEE Transactions on Multimedia, vol. 24, pp. 4433–4445, 2021

    2021

  20. [28]

    Skeleton-aware neural sign language translation,

    S. Gan, Y . Yin, Z. Jiang, and et al, “Skeleton-aware neural sign language translation,” in29th ACM Interna- tional Conference on Multimedia, 2021, pp. 4353–4361

    2021

  21. [29]

    Two-stream net- work for sign language recognition and translation,

    Y . Chen, R. Zuo, F. Wei, and et al, “Two-stream net- work for sign language recognition and translation,”Ad- vances in Neural Information Processing Systems, vol. 35, pp. 17043–17056, 2022

    2022

  22. [30]

    Ustm: Unified spa- tial and temporal modeling for continuous sign language recognition,

    A. A. Hasanaath and H. Luqman, “Ustm: Unified spa- tial and temporal modeling for continuous sign language recognition,”arXiv preprint arXiv:2512.13415, 2025

    2025

  23. [31]

    Openpose: Re- altime multi-person 2d pose estimation using part affin- ity fields,

    Z. Cao, G. Hidalgo, T. Simon, and et al, “Openpose: Re- altime multi-person 2d pose estimation using part affin- ity fields,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 1, pp. 172–186, 2019

    2019

  24. [32]

    Temporal convolutional networks for action segmentation and de- tection,

    C. Lea, M. D. Flynn, R. Vidal, and et al, “Temporal convolutional networks for action segmentation and de- tection,” 2016

    2016

  25. [33]

    Neural sign language translation,

    N. C. Camgoz, S. Hadfield, O. Koller, H. Ney, and R. Bowden, “Neural sign language translation,” inIEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2018

    2018

  26. [34]

    Improving sign language translation with monolingual data by sign back-translation,

    Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li, “Improving sign language translation with monolingual data by sign back-translation,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 1316–1325

    2021

  27. [35]

    Stochastic transformer networks with linear compet- ing units: Application to end-to-end sl translation,

    A. V oskou, K. P. Panousis, D. Kosmopoulos, and et al, “Stochastic transformer networks with linear compet- ing units: Application to end-to-end sl translation,” in IEEE/CVF International Conference on Computer Vi- sion, 2021, pp. 11946–11955

    2021

  28. [36]

    Mska: Multi- stream keypoint attention network for sign language recognition and translation,

    M. Guan, Y . Wang, G. Ma, and et al, “Mska: Multi- stream keypoint attention network for sign language recognition and translation,”Pattern Recognition, vol. 165, pp. 111602, 2025

    2025

  29. [37]

    Cross-modality data augmentation for end-to-end sign language translation. arxiv 2023,

    J. Ye, W. Jiao, X. Wang, and et al, “Cross-modality data augmentation for end-to-end sign language translation. arxiv 2023,”arXiv preprint arXiv:2305.11096

    2023 arXiv

  30. [38]

    Crosslingual generalization through multitask finetun- ing,

    N. Muennighoff, T. Wang, L. Sutawika, and et al, “Crosslingual generalization through multitask finetun- ing,” in61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2023, pp. 15991–16111

    2023

  31. [39]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, and et al, “Lora: Low-rank adaptation of large language models,” 2021

    2021

  32. [40]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, and et al, “Scaling instruction-finetuned language models,” 2022

    2022

  33. [41]

    Multilingual denois- ing pre-training for neural machine translation,

    Y . Liu, J. Gu, N. Goyal, and et al, “Multilingual denois- ing pre-training for neural machine translation,” 2020

    2020

  34. [42]

    Ararea- soner: Evaluating reasoning-based llms for arabic nlp,

    A. Hasanaath, A. Alansari, A. Ashraf, and et al, “Ararea- soner: Evaluating reasoning-based llms for arabic nlp,” arXiv preprint arXiv:2506.08768, 2025

    2025

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.