SACQ: Structured Decoding with Memory-Conditioned Refinement for Long-Horizon Forecasting

Name: SACQ: Structured Decoding with Memory-Conditioned Refinement for Long-Horizon Forecasting
Uploaded: 2026-10-15T05:01:08.329336+00:00
Duration: 930 s
Description: SACQ replaces simple flatten readouts in long-horizon forecasting with a two-stage decoding pipeline. It first creates a coarse patch scaffold, then refines it using memory-conditioned cross-attention, and finally fuses them with a learned gate. This explicit structure stabilizes training under long horizons and noisy labels.

SACQ replaces simple flatten readouts in long-horizon forecasting with a two-stage decoding pipeline. It first creates a coarse patch scaffold, then refines it using memory-conditioned cross-attention, and finally fuses them with a learned gate. This explicit structure stabilizes training under long horizons and noisy labels.

Open the full episode