BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate

Researchers introduced BARRED, a synthetic-data framework for training custom AI guardrail classifiers.

BARRED uses task decomposition plus debate-based verification to produce higher-fidelity policy labels.

Finetuned small language models trained on BARRED data outperformed proprietary LLMs and dedicated guardrail systems on several policies.

The approach could make task-specific safety filters cheaper, faster, and easier to scale than heavy human annotation.