acceptodds
Desk rejected at ICLR 2027

Rotors: Lightweight Geometric Recurrence for Edge Intelligence

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Applications of AI on edge devices are hampered by hardware limitations. To overcome these limitations, we re-frame sequence modeling through rotations on cooperating manifolds. Applied to sequence decoding, this enables stable recurrence over extended contexts that match standard Transformer baselines in perplexity with only memory complexity at inference. In bypassing the context bottleneck of Transformers, Rotors offer a mathematically robust, deterministic and performant paradigm for real-time edge intelligence.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Edge Intelligence, Sequence Modeling, Recurrent Neural Networks, Hyperspherical Manifolds, Householder Reflections, Constant-Memory Inference

Related papers

Discussion (0)

Sign in to comment.