Abstract
The computational cost of vision transformers (ViTs) can be a limitation for resource constrained systems. A predictive, dual-stream token routing system can dynamically route visual data tokens by combining top-down predictive information, such as a planned future trajectory, with a bottom-up, real-time analysis of the visual scene to detect anomalies. Based on this fused information, tokens can be selectively directed to a full-resolution processing pathway or a context-preserving compression pathway. This approach may reduce the computational load of ViT pipelines while supporting the preservation of information related to safety awareness, temporal stability, and overall scene context.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kligvasser, Idan; Kimhi, Moshe; Intrator, Yotam; Rivlin, Ehud; and Cohen, Regev, "Predictive and Safety-Aware Token Routing for Vision Transformer Pipelines", Technical Disclosure Commons, (October 12, 2026)
https://www.tdcommons.org/dpubs_series/12090