Training technique that separates prediction tokens from context encoding, allowing longer forecasts without masking observed values.