Token-Conditioned Poles SSM Improves Vision State Space Efficiency

A new paper on arXiv (2605.11563) introduces Token-Conditioned Poles SSM (TCP-SSM), a structured selective State Space Model framework for long-range vision tasks. TCP-SSM makes recurrence dynamics explicit and interpretable through stable poles, using real poles for monotone or sign-alternating decay and complex-conjugate poles for damped oscillatory responses. The approach improves efficiency over standard SSMs by controlling state-dependent memory behavior, particularly in compact backbones where long scan paths exceed effective memory horizon. The paper is a cross submission, likely from a computer vision or machine learning conference.