Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Geo-Sign: Hyperbolic Contrastive Regularisation for Geometrically Aware Sign Language Translation

Authors: Edward Fish, Richard Bowden

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the CSL-Daily benchmark [88] demonstrate Geo-Sign s efficacy. Our skeletal-based approach not only achieves a +1.81 BLEU4 and +3.03 ROUGE score over state-of-the-art pose-based methods but also matches the performance of comparable vision-based networks.
Researcher Affiliation Academia Edward Fish CVSSP, University of Surrey EMAIL Richard Bowden CVSSP, University of Surrey EMAIL
Pseudocode Yes Algorithm 1 Iterative Weighted Fréchet Mean in Bdhyp c
Open Source Code Yes Code available here: https://github.com/ed-fish/geo-sign.
Open Datasets Yes We evaluate Geo-Sign on Chinese Sign Language (CSL) and American Sign Language (ASL). For CSL we use the CSL-Daily dataset [88, 89], a large-scale corpus for Chinese Sign Language to Chinese text translation, comprising over 20,000 videos. For American Sign Language we perform translation experiments on How2Sign [15] and isolated sign language recognition experiments on WLASL2000 [39].
Dataset Splits Yes Experiments on the CSL-Daily benchmark [88] demonstrate Geo-Sign s efficacy... Ablation studies on the CSL-Daily test set for our best performing Geo-Sign (Hyperbolic Token) model are presented in Table 4. ... Data from CSL-Daily (Train: 18,401 sentences / 20.62 hours).
Hardware Specification Yes Hardware: 4 NVIDIA RTX 3090 GPUs.
Software Dependencies No Our implementation relies on Py Torch [56] as the primary deep learning framework. For Transformer models, we utilize the Hugging Face Transformers library. All hyperbolic geometry operations and Riemannian optimization are handled by the Geoopt library [36]. For distributed training and profiling, Deep Speed is employed.
Experiment Setup Yes Optimization employs Adam W [35, 45] for Euclidean parameters (ST-GCNs, m T5, linear layers), with learning rate 3 10 5. Hyperbolic parameters, including the learnable curvature c (optimized in log-space, e.g., log c) and manifold-constrained parameters, use Riemannian Adam (RAdam) [3] with a comparable learning rate. Key hyperparameters for the hyperbolic components (initial curvature c = 1.5, dimension dhyp = 256, and α = 0.70) are minimally tuned on the development set (further details in the appendix). In Table 7 we provide the full hyper-parameters for the best performing model.