In autoregressive generation, a model predicts the next step in a sequence based on its own previous outputs. This works beautifully for short bursts, but as sequences grow longer—stretching into minutes of video or weeks of weather patterns—a fundamental flaw emerges: exposure bias. Because the model is trained on perfect ground-truth data but must generate based on its own imperfect predictions, small errors compound. Eventually, the model "drifts" off the manifold of realistic data, leading to visual collapse, identity loss, or physical nonsense.
A growing body of research suggests that the solution is not to avoid errors, but to embrace them during training. By intentionally feeding a model its own flawed predictions, researchers are teaching systems to recognize drift and steer themselves back to stability.
The Teacher Forcing Trap
Standard training typically uses "teacher forcing," where the model is always given the correct previous frame or state. While efficient, this creates a "student-teacher mismatch." At inference time, the "teacher" is gone. If the model makes a 1% error in frame 10, it has never learned how to handle a 1% error in its input; it assumes the input is perfect. This leads to a feedback loop where errors amplify exponentially.
To break this cycle, Wuyang Li, Wentao Pan, and colleagues introduced Error-Recycling Fine-Tuning (ERFT) in Stable Video Infinity. Instead of shielding the model from its mistakes, ERFT forces the model to generate sequences based on its own noisy history. This teaches the Diffusion Transformer (DiT) to actively identify and correct artifacts rather than letting them accumulate.
Learning the Recovery Vector
The intuition behind this shift is that the model needs to learn a "recovery" behavior. If a generated video frame is slightly too dark or a motion stroke is slightly off-target, the model should learn to compensate in the next step.
In Self-Forcing++, Justin Cui and Jie Wu demonstrate that training on long-horizon, imperfect trajectories allows a "student" model to develop robustness against motion collapse. By observing how errors evolve over time, the model learns the underlying dynamics of the data manifold, allowing it to navigate back toward realistic states even after a deviation. Similarly, Shuo Chen and Cong Wei propose Context Forcing in Context Forcing: Consistent Autoregressive Video Generation with Long Context, which uses a teacher that understands long-term history to guide a student through these recycled errors.
Mechanisms of Self-Correction
Researchers have converged on several key mechanisms to implement this self-correction:
- Replay Buffers: In the weather forecasting system FengWu, Kang Chen and LEI BAI use a replay buffer to store and sample previous predictions. This allows the model to "practice" on its own forecasts, which is essential for extending skillful global predictions beyond the traditional 10-day limit.
- Scheduled Sampling: In DART, Kaifeng Zhao and Gen Li use a schedule that gradually replaces ground-truth data with the model's own predictions. This "warm-up" period prevents the model from being overwhelmed by noise early in training while ensuring it is robust by the end.
- Recursive Feedback: For specialized tasks like video colorization, Markus Hofinger and Thomas Pock show in Learned Variational Video Color Propagation that feeding erroneous outputs back into an unrolled optimization loop helps the network learn to mitigate temporal drift.
Stability Across Domains
This paradigm shift is proving effective across diverse modalities. In the realm of digital humans, SoulX-LiveTalk by Shunshun Yin uses Self-Correcting Bidirectional Distillation to prevent identity drift during infinite-duration streaming. In autonomous driving, Epona by Kaiwen Zhang employs a "chain-of-forward" strategy to ensure that simulations of the road remain consistent for minutes at a time.
The shared insight across these works—from Painting Many Pasts to WorldCompass—is that long-term stability is not a byproduct of higher accuracy, but a learned capability. By treating its own errors as training data, a model learns not just to predict the future, but to maintain it.