Apple presents parallel RNN training, improved state space models, and unified vision models at ICLR 2026

Apple researchers are presenting five major research contributions at the Fourteenth International Conference on Learning Representations (ICLR 2026) in Rio de Janeiro this week, advancing work in efficient sequence modeling, multimodal vision-language systems, and 3D generation.

The centerpiece of Apple's presentation is ParaRNN, a framework for parallelizing training of recurrent neural networks. Historically, RNNs have been efficient for inference but difficult to scale due to the sequential nature of their computation. Apple's new approach achieves a 665× speedup over traditional sequential RNN training, making it feasible to train classical RNNs with up to 7 billion parameters. Benchmarking shows these large-scale RNNs achieve language modeling performance competitive with transformers of comparable size, potentially opening new architectural choices for practitioners building models under computational constraints. The ParaRNN codebase has been released as open source.

In parallel work on state space models (SSMs), Apple researchers identify and address a fundamental limitation: SSMs excel at long-context inference due to fixed-size memory and linear computational scaling, but this same constraint prevents them from solving complex problems that exceed the model's capacity, even with chain-of-thought generation. The research demonstrates that providing SSMs with external tool access—such as memory tools or code execution—enables them to generalize to arbitrary problem length and complexity on arithmetic, reasoning, and coding tasks.

Apple's Manzano model tackles a design trade-off in unified vision-language systems. Many existing multimodal models must choose between strong image understanding or strong generation. Manzano uses a hybrid tokenizer with separate lightweight adapters that feed a shared semantic space: one adapter produces continuous embeddings for understanding, the other produces discrete tokens for generation. A unified autoregressive language model predicts both text and image tokens, with an auxiliary diffusion decoder converting image tokens to pixels. This design achieves state-of-the-art results among unified models while remaining competitive with specialist systems, particularly on text-heavy evaluation.

Apple is also demonstrating local LLM inference on Apple silicon using the MLX framework and techniques for fast 3D scene synthesis from single images. The company is sponsoring affinity group events supporting underrepresented groups in machine learning research.