Apple's Machine Learning Research at ICLR 2026

Apple will present a wide range of research at ICLR 2026, spanning RNNs, SSMs, multimodal generation, 3D scene synthesis, and protein folding. The event will be held in Rio de Janeiro, Brazil, with Apple also participating as a sponsor.

One of the key presentations is ParaRNN, a framework that parallelizes the training of nonlinear RNNs, achieving 665x faster training compared to traditional sequential methods. This approach makes it possible to train classic RNNs at the 7B parameter scale, allowing them to potentially compete with transformers in language model performance, and the code has also been open-sourced.

Another paper, To Infinity and Beyond: Tool-Use Unlocks Length Generalization in State Space Models, examines the strengths and weaknesses of SSMs. Even architectures like the Mamba family, which excel at long context, show degraded performance when generating text longer than the model's capacity. However, the paper demonstrates that giving the model access to external tools, combined with appropriate training data, enables generalization to greater lengths and higher complexity on arithmetic, reasoning, and coding tasks.

In the area of handling images and text together, MANZANO is introduced. A single shared vision encoder produces continuous embeddings for understanding and discrete tokens for generation, and a unified autoregressive LLM paired with an auxiliary diffusion decoder handles both image understanding and generation based on this. Apple explains that this architecture and training recipe reduces the trade-off between understanding and generation performance found in existing open-source unified models.

Results are also emerging in the 3D and biology fields. SHARP generates a 3D Gaussian representation from a single photo in under 1 second, with real-time rendering possible from nearby viewpoints. It provides a metric representation with absolute scale, achieving 25-34% LPIPS and 21-43% DISTS improvements over existing top models, while reducing synthesis time by three orders of magnitude. SimpleFold presents an approach to solving protein folding using only standard transformer blocks instead of complex specialized architectures, and provides code and checkpoints, enabling local execution on Apple silicon with MLX.

At booth #204, visitors can experience an SHARP demo alongside hands-on MLX-based local LLM inference. Apple is also sponsoring and participating in affinity group events such as Women in Machine Learning (WiML) and Queer in AI, continuing to combine research presentations with community support on-site at ICLR.