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

arXiv · Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Lida Zhao, Yutao Wu, Simin Chen, Ying Zhang, Leo Yu Zhang · 2026-09-27

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Jev turns a single LLM call into a calibrated, zero-shot detector for ten common alignment failures.

What it does

The authors introduce Jev, an RL-trained model that answers a wide range of alignment-related questions with calibrated probabilities in one inference step. They benchmark Jev on RLCDAlignBench, covering ten failure modes (sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, uncertainty concealment, power seeking) across 44 benchmarks and five target models. A generic question achieves a median AUROC of 0.886 zero-shot, out-performing supervised baselines on most tasks, while question wording has little effect and input context matters more. Jev matches human-aligned scorer agreement and runs about 63× cheaper than typical LLM-judge pipelines.

Why it matters

Alignment monitoring is a massive cost center; a single, calibrated query that reliably flags multiple failure modes can slash monitoring budgets and simplify pipelines. The high AUROC demonstrates that zero-shot detection is viable, reducing the need for bespoke classifiers per failure type.

How it applies

Teams can replace multi-step judge calls with a single Jev query that returns calibrated risk scores for each failure mode, integrating the result directly into guardrails or automated remediation. Because context fields drive performance, practitioners can craft minimal input wrappers to surface the relevant label cues without retraining.

The short version

One query, ten alignment red-flags—Jev makes guardrails cheap and fast.

My take

From a builder’s standpoint, Jev is a template for turning RL-calibrated models into universal auditors. The fact that a generic prompt works so well means you can roll this out across models without heavy prompt engineering. The 63× cost advantage alone makes it a no-brainer for any operation that runs alignment checks at scale.

How it connects

• Aligns with the trend of moving from per-failure classifiers to unified, multi-task auditors.
• Shows the practical upside of RL-based calibration beyond pure reward maximization.
• Highlights the growing need for cost-effective, zero-shot safety tooling as LLMs proliferate.

Bottom line

Swap your multi-step alignment judges for a single Jev call and cut monitoring costs while keeping detection quality high.

Takeaways

• Jev uses RL-trained calibrated decisions to answer many alignment-failure questions in one forward pass.
• On RLCDAlignBench, a generic prompt reaches median AUROC 0.886 zero-shot, beating most supervised baselines.
• Question wording is largely irrelevant; the input context (field encoding) drives detection quality.
• Jev’s inference cost is roughly 1/63 of typical LLM-judge pipelines.
• The benchmark reveals label inconsistencies in existing alignment datasets.

Topics: alignment, zero-shot detection, RL fine-tuning, calibrated probabilities, evaluation