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

Just Ask Jev. Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures. By Ruoqi Guo and eight others, from Griffith University, Nanyang Technological University, and five other institutions. Published on the twenty-fourth of September. Its code repository is public and real: about twenty-five source files.

Detectors of alignment failures do two jobs. They screen language models that have been deployed, and they score alignment benchmarks. Most of them are generative judges, which spend a full decoding pass on every criterion they check. Classifiers that read token probabilities, Llama Guard being the example the abstract gives, avoid that decoding, but still score one fixed label each time. Jev is a different kind of model. It was trained with reinforcement learning for calibrated decisions, or RLCD, and it answers many typed questions about a single input with calibrated probabilities, all in one call. Whether that ability actually detects alignment failures had never been measured.

RLCDAlignBench is the measurement. It benchmarks Jev on ten alignment failures: sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking. The whole thing spans forty-four benchmarks and five target models. Labels come from each benchmark's own scorer, and on two of the benchmarks, from humans as well.

One observation drives the design. Many of these failures are relational. They are defined against a reference, such as what the user believes or an instruction that was injected, and the response on its own does not reveal that reference. A detector that reads only the response is missing half of the case. So the authors vary two things separately. On one side is what Jev is asked: the wording of the question, and the type of answer it expects. On the other side is what Jev sees: which fields of the input it is given.

A single generic question, used zero-shot, reaches a median AUROC of zero point eight eight six, and beats supervised baselines on most benchmarks. The two axes of the design come apart cleanly. Question wording matters little. Context matters more, and mostly through fields that encode the label.

Against people, Jev matches the reference scorer's agreement with human labels. Along the way it surfaces label defects in existing benchmarks. And it costs sixty-three times less than scorers built on an LLM judge.

The title's advice turns out to be literal. Just ask: the wording of the question barely moves the result, and one generic question already does most of the work. The lever is on the other side of the split, in which fields of the input Jev is shown. For a detector of failures that are defined against a reference, what it sees matters more than how it is asked.