Researchers fine-tune SHuBERT-ByT5 for real-time sentence-level sign language translation
A team fine-tuned a SHuBERT-ByT5 translation stack on a subset of How2Sign using QLoRA, achieving a BLEU of 15.9 on the test set. The work focuses on real-time deployment rather than novel architectures, addressing the gap between isolated sign recognition and natural sentence-level translation.
Entities
Related
VTaMo introduces explicit multi-granularity alignment for sign language translation
VTaMo is a new framework for sign language translation that introduces explicit multi-granularity alignment between video and text at three levels: local alignment via optimal transport with a learnable null token, global alignment via orthogonal transformation, and sequence-level alignment. This approach aims to improve gloss-free SLT by providing more structured cross-modal correspondences.
Tokenizer Transplantation Fixes Autoregressive Collapse in Bengali ASR
Researchers propose a tokenizer transplantation pipeline to replace the English-centric byte-level tokenizer in Moonshine speech recognition models with the native-script BanglaBERT WordPiece vocabulary. This resolves catastrophic autoregressive collapse on Bengali, enabling edge-efficient ASR for morphologically rich non-Latin languages.
Researchers identify modality interference as root cause of full-duplex SLM degradation
A new paper on arXiv (July 7, 2026) presents a fine-grained analysis of optimization dynamics in full-duplex Spoken Language Models (SLMs), identifying severe modality interference as the root cause of knowledge degradation and compromised semantic integrity. The work aims to enable more natural and intelligent full-duplex SLMs.
Paper challenges text-only pretraining, proposes visual pretraining for language models
A new arXiv paper argues that current language model pretraining discards rich visual information from documents and web pages. The authors propose scalable visual pretraining to incorporate figures, equations, and layouts, aiming to improve language intelligence beyond text-only approaches.
Soofi S 30B-A3B: Open-source MoE hybrid Mamba Transformer for German and English
Researchers released Soofi S 30B-A3B, a sovereign open-source Mixture-of-Experts foundation model for German and English. Its hybrid Mamba-Transformer design activates only 3B of 30B parameters per token, achieving throughput advantages for long-context deployment. Pretrained on 27 trillion tokens with up-weighted German data, it matches dense 14-27B models on English and German benchmarks while excelling in code tasks.