SignLLM

SignLLM is a multilingual Sign Language Production (SLP) large language model that generates sign language gestures from text inputs. It introduces two novel SLP modes: MLSF (Multilingual Sign Language Framework) and Prompt2LangGloss, enabling sign language production from query texts and question-style prompts. The model incorporates a reinforcement learning loss and a Priority Learning Channel module to accelerate training by improving data sampling. For training, a comprehensive multilingual dataset called Prompt2Sign is built from public data covering American Sign Language (ASL) and seven other sign languages (DGS, KSL, DSGS, LSF-CH, LIS-CH, LSA, TSL). SignLLM achieves state-of-the-art performance across eight sign languages and is developed in collaboration with commercial companies (not open-sourced).