Apple AI Updates: September 12, 2026

1. DiscoSign Moves Sign Language Translation Above the Sentence Level

Apple published DiscoSign, a modular LLM-based framework for text to sign language gloss translation that handles discourse phenomena sentence-level systems drop. It addresses spatial coreference resolution, where entities have to keep consistent spatial locations across a whole discourse, along with Question-Answer Clauses, the pseudocleft structures that carry much of the information structure in signed languages. Source

2. Apple Proposes Scoring Video Captions by Multiple-Choice QA Instead of Reference Matching

A second paper argues that reference-matching metrics penalize good video captions for lexical mismatches or for focusing on a different but valid part of the frame, a consequence of the one-to-many nature of video description. The proposed alternative redefines caption quality as information fidelity and evaluates it by asking multiple-choice questions the caption should be sufficient to answer, which also yields a more fine-grained breakdown than a single similarity score. Source

3. SimpleDesign Collapses Protein Codesign Into a Single Training Stage

SimpleDesign is a joint generative model over protein sequence and 3D structure that drops the usual two-stage recipe, where autoencoders are first trained to tokenize data into latents and a generative model is then trained on those latents. Folding it into one stage removes the dependency on separately trained tokenizers, which is the part of the pipeline that tends to bound downstream quality in drug discovery and protein engineering work. Source