An arXiv preprint describes SignFLIP, a model built to handle sign-language translation and sign-language generation in one system. On pose-only translation tests, it reported BLEU-4 scores of 26.10 on CSL-Daily, 15.1 on How2Sign and 19.77 on OpenASL. The results sit alongside a more sobering signal: in a small human rating exercise, generated movements scored below the ground-truth motions used for comparison.
The paper describes a symmetric architecture and stage-wise training in which a shared sign-text representation is progressively refined across translation and generation. In other words, the proposed system is built around a common representation for the two tasks.
A broad test across sign-language data
The experiments used sign-language pose and text data from American Sign Language and Chinese Sign Language sources. The listed datasets include roughly 610,000 YouTube-ASL samples, 33,000 How2Sign samples, 97,000 OpenASL samples, 2,038 WLASL samples, roughly 700,000 CSL-News samples and 20,000 CSL-Daily samples.
Translation and generation scores
On CSL-Daily, pose-only SignFLIP recorded test BLEU-4 of 26.10 and a ROUGE score of 55.24. On How2Sign, it reported BLEU-4 15.1 and BLEURT 49.8; on OpenASL, BLEU-4 was 19.77 and BLEURT 59.35. These figures are translation results from pose data alone, under the dataset and input conditions tested.
For sign-language generation, the reported How2Sign row lists DTW 0.248, DTW-MJE 5.28e-4 and back-translation BLEU-4 7.10. On CSL-Daily, the corresponding figures are DTW 0.025, DTW-MJE 6.15e-5 and BLEU-4 9.75. The paper reports these as pose-generation and back-translation measures, alongside the translation scores.
The human check tells a different story
The human check was small: two professional signers rated 15 generated samples on a 1-to-5 scale. SignFLIP received average scores of 1.69 on How2Sign and 1.94 on CSL-Daily. Ground-truth motions scored 3.48 and 3.44, respectively. The gap is important because the authors say low ratings, including for ground-truth motions, raise concerns about how understandable 2D pose representations are for sign-language users.
Recognition transfer was also tested
SignFLIP's learned representations were also tested for sign-language recognition with the encoders frozen. The reported test word error rate was 28.8 on CSL-Daily. On WLASL100, top-1 accuracy was 88.87 in the P-I setting and 89.16 in P-C. These recognition results do not replace the human assessment of generated movement.
What changed when training stages were compared
Separate ablation comparisons reported gains linked to the training setup. The pretraining setting was associated with gains of 3.37 BLEU-4 on CSL-News and 4.26 on YouTube-ASL. For generation, the ablation text reported DTW improvements of 11% for pretraining and 16% for SLT training. The paper presents these as within-study comparisons.
The unresolved question is comprehension
The benchmark results do not resolve the human question raised by the paper's limitations. Many listed datasets are not primarily based on Deaf and hard-of-hearing signers, and CSL-Daily was produced by interpreters using source text. The authors also question whether the 2D outputs are sufficiently understandable for sign-language users.
The work is an arXiv preprint, version 1, dated 28 Sep 2026. Its reported evidence remains centered on computational benchmarks, ablations, recognition tests and a small human rating exercise.
Paper data and sources
Original title: SignFLIP: A Unified Model for Sign Language Translation and Generation via Stage-wise Alignment at Scale
Authors: Zhaoyi An, Sihan Tan, Youngbae Hwang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text