Distribution Shift

Distribution shift refers to the statistical distribution difference between the training environment and the deployment environment, and is the direct mechanism by which spurious correlations cause performance collapse.

Distribution shift refers to the systematic difference between the statistical distribution of data seen during model training and the distribution of data in the actual deployment environment.

The core challenge of offline RL is distributional shift, which can lead to extrapolation errors that cause the learned policy to perform far worse than expected at deployment.

Domain shift refers to the systematic difference between the distribution of training data and the distribution of data in actual deployment scenarios, and is one of the core challenges in transfer learning and AI engineering deployment

A model that achieves high mAP on academic benchmarks may perform far worse in production if there is a distribution gap between its training data and real-world scenarios, or if evaluation fails to account for practical factors such as camera noise and occlusion

Distribution shift can cause significant drops in model performance or even complete failure

Zero-shot models rely on visual-semantic associations learned during pre-training, and may make misjudgments when distractors similar to the training distribution appear in the scene.

Means of mitigating distribution shift include rolling backtests on recent data, introducing exogenous variables, and online calibration of model outputs; the fundamental measure is to establish monitoring that triggers human review or model updates when the distribution of forecast residuals changes significantly.

Even with the correct context length, time-series foundation models still struggle with distribution shift: if the training data does not contain future states, the model will interpolate confidently but may be wildly wrong; a longer context window cannot make up for this flaw.

Distribution Shift refers to the fact that the behavior an Agent exhibits during training and controlled testing may not represent its behavior in a real deployment environment.

When a neural network trained offline is deployed to a closed-loop control system, its actions change the system's state distribution, causing the data distribution at actual runtime to deviate from the offline data distribution at training time, that is, distribution shift.

Models based on causal mechanisms capture stable structures in the data-generating process and are theoretically more robust to distributional shifts.

The positive correlation between base task performance and generalization ability weakens under significant distribution shift.

GAP-DPO mitigates distribution shift caused by off-policy sampling through an epoch-wise regeneration mechanism that regenerates samples at each epoch.

Even with palm-size normalization, differences in hand structure, pinching habits, and finger-length ratios across individuals mean a single static threshold cannot simultaneously work for multiple users, which is fundamentally a distribution shift problem.

The real world is non-stationary, and a policy's performance during training may deteriorate sharply after the environment undergoes a distribution shift.

数据增强的核心原则是增强分布应当匹配真实测试分布，理想策略是让训练变换分布成为真实测试分布的一个超集