Apple's SimpleDesign ditches the latent space for protein co-design
Apple's ML research group has had a quietly productive week, and today it landed three papers at once: a protein-design model trained straight in the raw data space, a benchmark that grades video captions with quizzes instead of reference text, and a sign-language translator that finally pays attention to what was said two sentences ago. None is a consumer launch, but together they sketch where Apple's research muscle is pointed — biology, evaluation, accessibility.
Apple's SimpleDesign generates protein sequences and their 3D structures together, skipping the latent-space stage other co-design models rely on. Generative protein models typically train in two stages — first compress sequence and structure data into latent tokens with autoencoders, then train a generative model on those compressed representations. Apple's researchers bet the two-stage pipeline is unnecessary: SimpleDesign trains end-to-end directly in the data space, combining a discrete cross-entropy objective for the amino-acid sequence with a regression objective for the 3D structure, on a multimodal Transformer backbone that processes each modality differently but lets global self-attention span both. Trained on more than 2 million sequence-structure pairs, it posts competitive results across co-design benchmarks as well as unconditional generation of sequences and structures on their own. The significance is architectural honesty: after SimpleFold argued that AlphaFold-style architecture wasn't necessary for folding, this paper makes the same case for design — simpler pipelines, fewer stages, fewer places for training to go wrong. Drug-discovery and protein-engineering teams churning through design candidates should take note. We covered the upside of AI-designed proteins earlier this summer — an AI-designed protein delivers RNA better than lipid nanoparticles — and better generative tooling is exactly how that pipeline gets faster. Paper: arXiv:2609.03377.
Apple's CapQuiz grades video captions by quizzing models on them instead of comparing text against a reference. Video-caption evaluation has a known flaw: because many valid captions can describe the same clip, reference-based metrics penalize good captions for harmless wording differences, and they give one opaque score where you wanted a diagnosis. CapQuiz flips the setup — a caption is judged by how well a question-answering system performs on human-verified, fine-grained multiple-choice questions generated from the video, so a caption is good if the information in it survives the quiz. The benchmark spans 10 question types across 24 video domains and introduces CapF1, a composite of CapP (factuality) and CapR (coverage). Apple reports it correlates better with human judgment than existing metrics. Better caption metrics feed directly into better video models, since captions are the training data for text-to-video systems — sloppy scoring, sloppy models. Paper: arXiv:2609.09973, also published at ACL 2026.
Apple's DiscoSign makes text-to-sign-language translation discourse-aware, not just sentence-by-sentence. Sign language translation systems almost always work one sentence at a time, which breaks three things that matter in real conversation: entities should stay anchored to consistent spatial locations across a discourse, sign languages use pseudocleft Question-Answer Clause structures that sentence-level models can't represent, and the mapping between English concepts and ASL signs needs to stay consistent rather than drifting. DiscoSign, an LLM-based framework built on linguistic research, handles all three and ships a matching suite of evaluation metrics for discourse coherence — the first systematic framework for discourse-level gloss translation. Experiments show better spatial consistency and entity tracking than sentence-only translation while keeping single-sentence quality competitive. Accessibility AI keeps getting the research seriousness it deserves. Paper: arXiv:2609.02796.
What to watch: whether SimpleDesign's single-stage recipe gets adopted by the open-protein-design community — if a simpler pipeline wins, two-stage latent co-design may go the way of two-stage diffusion.
Three papers, one day — is Apple quietly building the strongest applied ML research bench in big tech? Tell us in the comments.
Sources: Apple Machine Learning Research · arXiv:2609.03377 · arXiv:2609.09973 · ACL Anthology · arXiv:2609.02796
