Chimera Training: Detecting logical anomalies without real-world rule violations

Logical anomaly detection traditionally struggles because real-world rule violations are rare during training. A new paper from researchers proposes Chimera Training, a technique that synthesizes counterfactual examples at the feature level. The model compiles each semantic constraint (e.g., "a cup must be on a table") into a directed acyclic graph, where each node represents a logical operator with an MLP gate that learns to map child features and negations to a parent representation and satisfaction probability. Intermediate supervision comes from exact Boolean propagation over ground-truth concept labels.

To overcome the lack of informative counterexamples in training data, Chimera Training constructs hybrid samples by concatenating subtree features from different images. Each operand retains the hard truth label from its source, and the target is computed by applying the node's logical operator to those inherited labels. This generates supervised logical counterexamples without requiring real anomalous images. Evaluated on CLEVRER, OpenImages, and VidOR, the method consistently outperforms independent-events and same-image semantic-training baselines in rule-level anomaly AUROC, particularly for compositional and relational rules. The system outputs both scalar anomaly scores and per-rule attributions, enabling interpretable detection.