New method fixes unreliable training signals in long-range AI forecasting

Original: Large Distant Gradients Need Not Be Reliable: reliability-weighted credit assignment for long-horizon autoregressive forecasting

Source: arxiv.org ↗