Progressive Information Distillation Transformer for Channel Estimation in RIS-Assisted Systems

Reconfigurable Intelligent Surfaces are pivotal for 6G millimeter-wave systemsCyet their passive and large-scale nature imposes excessive pilot overhead in cascaded channel estimation. To address this, we propose a Progressive Information Distillation Transformer for channel reconstruction from low-dimensional observations via multi-stage feature distillation. The core innovation lies in an iterative dual-stream architecture: a preservation stream retains critical low-redundancy features, while a refinement stream aggressively processes redundant details via a synergistic Local Window and Global Grid Attention mechanism, effectively capturing complex spatial dependencies. Simulations demonstrate the scheme’s superior performance, achieving significant gains in channel estimation accuracy with a compact parameter footprint.