Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

Summary

The paper introduces RLCDAlignBench, an evaluation of Jev, a commercial model trained with reinforcement learning for calibrated decisions (RLCD) that returns probabilities for many typed questions in one API call, as a zero-shot detector of ten alignment failure types. The benchmark spans 44 alignment benchmarks, 7,193 detection instances and five open 2-7B target models, labelled by each benchmark's reference scorer and, on two benchmarks, by humans. The design separates what the detector is asked (generic template versus targeted wording; three answer types) from what it sees (state fields added one at a time), with a split-half protocol that prevents selection inflation. A single generic question reaches a median AUROC of 0.886 zero-shot and beats TF-IDF and length baselines on 25 of 31 benchmarks; targeted wording adds only +0.006 AUROC out of sample. On StrongREJECT, Jev matches the GPT-4o-mini scorer's agreement with humans (kappa 0.809 versus 0.811) and ranks better (0.971 versus 0.929), and its confident disagreements exposed label defects in several benchmarks, at 63 times lower cost than LLM judges. Limitations: per-benchmark miscalibration (median ECE 0.168) forces threshold fitting on about ten labels; most labels are unvalidated scorer outputs, raising judge-circularity concerns; only one RLCD model, English data and small targets are covered.