teacher forcing mitigation

Training strategies that reduce mismatch from teacher forcing and chunk-wise autoregressive denoising by approximating inference-time autoregressive feedback (rollouts) without slow sequential sampling, enabling parallelized training while preserving effective sequence length.