Worcester Polytechnic Institute

The work proposes GSL-Diffusion, a time-series-conditioned diffusion channel estimator that denoises in the angular domain: a bidirectional LSTM encodes a short sequence of least-squares observations and injects global and sequential conditions into a 3D U-Net via multi-scale cross-attention, while a learnable SNR-gated late-fusion shortcut adaptively trades observation fidelity against the generative prior, and SNR-adaptive truncated DDIM inference reduces reverse steps; on QuaDRiGa/3GPP 38.901 time-varying channels it achieves the best NMSE almost across roughly -10 to 20 dB versus LS, LMMSE and diffusion baselines, often reaching one reverse step at high SNR.

The work proposes GSL-Diffusion, a time-series-conditioned diffusion channel estimator that denoises in the angular domain: a bidirectional LSTM encodes a short sequence of least-squares observations and injects global and sequential conditions into a 3D U-Net via multi-scale cross-attention, while a learnable SNR-gated late-fusion shortcut adaptively trades observation fidelity against the generative prior, and SNR-adaptive truncated DDIM inference reduces reverse steps; on QuaDRiGa/3GPP 38.901 time-varying channels it achieves the best NMSE almost across roughly -10 to 20 dB versus LS, LMMSE and diffusion baselines, often reaching one reverse step at high SNR.

This work demonstrates on a live O-RAN testbed that two autonomous agents with individually correct objectives—one protecting a latency SLA and one maximizing utilization for energy efficiency—jointly drive recurring opposing excursions of the shared resource partition, and presents and proves a lightweight arbitration layer, AURA, whose three admission checks (feasibility invariants, per-variable dwell time, deadband) reduce shared-state excursion amplitude from 8.4 to 0.4 PRB and cross-slice throughput starvation from 40–55% to 0.3%, while leaving the protected slice's own latency compliance unchanged.

This work demonstrates on a live O-RAN testbed that two autonomous agents with individually correct objectives—one protecting a latency SLA and one maximizing utilization for energy efficiency—jointly drive recurring opposing excursions of the shared resource partition, and presents and proves a lightweight arbitration layer, AURA, whose three admission checks (feasibility invariants, per-variable dwell time, deadband) reduce shared-state excursion amplitude from 8.4 to 0.4 PRB and cross-slice throughput starvation from 40–55% to 0.3%, while leaving the protected slice's own latency compliance unchanged.