Out-of-Distribution

Out-of-Distribution (OOD) detection is a key research direction in the field of machine learning, which focuses on identifying data samples that were not covered during the model training phase. This detection technique is crucial to improving the robustness of the model, especially when the model may encounter new environments that are significantly different from the training data. The core challenge of OOD detection is that the model needs to be able to respond correctly when faced with unknown or abnormal data, rather than blindly making predictions. These unknown samples may come from completely different distributions or have different characteristics from the training data, which requires the model to have a certain generalization ability.

In practical applications, OOD detection has a wide range of application scenarios, including but not limited to medical diagnosis, financial risk assessment, autonomous driving, etc. In these fields, the model's decision may have a significant impact, so it is particularly important to be able to accurately identify and handle OOD samples. For example, in medical diagnosis, the model may encounter rare cases that may never appear in the training data. If these OOD samples cannot be correctly identified, it may lead to incorrect diagnosis.

Shanghai Jiao Tong University and Alibaba Tongyi Laboratory published a paper in 2024 titled "Embedding Trajectory for Out-of-Distribution Detection in Mathematical Reasoning", which has been accepted by NeurIPS 2024, is the first research result on out-of-distribution detection in mathematical reasoning scenarios.