From the source
LG AI Research presented an ICML 2026 paper on label shift in tabular foundation models, showing that TabPFN's minority-class recall collapses under class imbalance.
- The paper, "Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment," proposes a simple adjustment at inference time rather than retraining.
- TabPFN is pre-trained on large-scale synthetic data and predicts in a single forward pass with no parameter updates, an in-context learning approach that has drawn attention for beating XGBoost on small datasets.
- The authors found that TabPFN folds the class imbalance present in its context directly into its predictions: in extreme cases 98.3% of test samples were assigned to the majority class.
- An analysis of 253 OpenML tabular datasets found roughly 85%, or 214, showed class imbalance, a pattern the authors say is common in defect detection, anomaly detection and disease diagnosis.
- The problem appeared even when training and testing used the same dataset, which the authors say underscores its severity.
- LG AI Research attended ICML 2026 at COEX in Seoul from July 6 to 11, 2026, the conference's 43rd edition and the first held in South Korea, where 23,918 papers were submitted and 6,352 accepted.
