Adversarial Data Curation via Failure-Mode Contextual Bandits
August 20, 2026
This framework uses contextual bandits to automate the selection of synthetic training data by targeting specific model failure modes. By clustering failures and using an LLM judge ensemble, the method improved RoBERTa-base accuracy on SNLI from 88.48% to 92.60%.
HOW THIS AFFECTS YOU
●
builderThis provides a path toward automated, self-improving data pipelines for niche NLU tasks.
●
researcherYou can use this adaptive sampling approach to improve model robustness more efficiently than fixed-threshold synthetic data generation.