Researchers debut SABRE for trustworthy AI out-of-distribution detection
A trio of AI agents now automatically picks the right anomaly detector on the fly.
Researchers from the University of Michigan introduced SABRE (Selective Agentic Budgeted Reliability Ensemble), a multi-agent framework that replaces static post-hoc OOD detectors with dynamic, per-regime selection at inference time. The system deploys three cooperative agents: a Selector that chooses which detector to query next, a Reporter that consolidates anomaly scores, and an Analyst that calibrates detector reliability using a small, labeled holdout sample unseen during training. This trio operates under a strict query budget and never observes test labels, ensuring deployment-time adaptability without benchmark assumptions.
In experiments, SABRE consistently outperforms fixed-detector baselines by avoiding ‘detector inversion’—where a detector trained on one domain misclassifies in-distribution data as more anomalous than genuine outliers in another. The framework’s Analyst component was particularly decisive, ruling out unreliable detectors and preventing score aggregation from canceling out sound signals. The authors also propose four new multimodal density detectors included in SABRE’s library, demonstrating that agent-driven selection yields reliable OOD detection across diverse domains without prior knowledge of optimal configurations.
- SABRE uses three specialized agents (Selector, Reporter, Analyst) to dynamically choose OOD detectors under a query budget
- Analyst calibration prevents detector inversion by weighting reliability without test labels, improving robustness across domains
- New multimodal density detectors and agent feedback deliver 15–30% gains in OOD detection accuracy vs. fixed baselines
Why It Matters
SABRE enables trustworthy AI systems to self-correct for domain shift, reducing costly model failures in production.