SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting

Name: SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting
Uploaded: 2026-10-15T04:52:02.073305+00:00
Duration: 816 s
Description: SteerCast improves decoder-only time series forecasting during inference without retraining parameters. It works by creating a database of historical representations and steering vectors derived from model errors. At test time, it retrieves similar past scenarios and injects correction signals into the model's hidden states to guide predictions toward ground-truth trajectories.

SteerCast improves decoder-only time series forecasting during inference without retraining parameters. It works by creating a database of historical representations and steering vectors derived from model errors. At test time, it retrieves similar past scenarios and injects correction signals into the model's hidden states to guide predictions toward ground-truth trajectories.

Open the full episode