Sihan Tan

Unifying sign translation and generation within a single LLM reveals that representations adapted for translation directly boost generative fidelity, matching bespoke task-specific architectures across the board.