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

Hugging Face Daily Papers Papers
ai-alignment reinforcement-learning zero-shot-detection benchmark safety

Summary

This paper presents RLCDAlignBench, a benchmark for evaluating Jev, a model trained with reinforcement learning for calibrated decisions, as a zero-shot detector for ten AI alignment failures with high performance and cost efficiency.

Detectors of alignment failures screen deployed language models and score alignment benchmarks. Most are generative judges that spend a decoding pass on every criterion, and classifiers that read token probabilities, such as Llama Guard, still score one fixed label per call. Jev, a model trained with reinforcement learning for calibrated decisions (RLCD), answers many typed questions about one input with calibrated probabilities in a single call. Whether it detects alignment failures has not been measured. We present RLCDAlignBench, which benchmarks Jev on ten alignment failures: sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking. It spans 44 benchmarks and five target models, labelled by each benchmark's scorer and, on two, by humans. Many of these failures are relational, defined against a reference, such as the user's belief or an injected instruction, that the response alone does not reveal. Our key idea is therefore to vary what Jev is asked separately from what it sees: the question's wording and answer type on one side, the fields of the input on the other. A single generic question reaches a median AUROC of 0.886 zero-shot and beats supervised baselines on most benchmarks. Question wording matters little, while context matters more, mostly through fields that encode the label. Jev matches the reference scorer's agreement with human labels, surfaces label defects in existing benchmarks, and costs 63x less than LLM-judge scorers. Code and data: https://github.com/sumleo/RLCDAlignBench.
Original Article
View Cached Full Text

Cached at: 09/25/26, 11:45 AM

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

Source: https://huggingface.co/papers/2609.29429 Published on Sep 24

·

Submitted byhttps://huggingface.co/sumleo

YI LIUon Sep 25

Abstract

Detectorsofalignmentfailuresscreendeployedlanguagemodelsandscorealignmentbenchmarks.Mostaregenerativejudgesthatspendadecodingpassoneverycriterion,andclassifiersthatreadtokenprobabilities,suchasLlamaGuard,stillscoreonefixedlabelpercall.Jev,amodeltrainedwithreinforcementlearningforcalibrateddecisions(RLCD),answersmanytypedquestionsaboutoneinputwithcalibratedprobabilitiesinasinglecall.Whetheritdetectsalignmentfailureshasnotbeenmeasured.WepresentRLCDAlignBench,whichbenchmarksJevontenalignmentfailures:sycophancy,jailbreaks,deception,promptinjection,hallucination,privacyviolation,socialbias,rewardhacking,concealinguncertainty,andpowerseeking.Itspans44benchmarksandfivetargetmodels,labelledbyeachbenchmark’sscorerand,ontwo,byhumans.Manyofthesefailuresarerelational,definedagainstareference,suchastheuser’sbelieforaninjectedinstruction,thattheresponsealonedoesnotreveal.OurkeyideaisthereforetovarywhatJevisaskedseparatelyfromwhatitsees:thequestion’swordingandanswertypeononeside,thefieldsoftheinputontheother.AsinglegenericquestionreachesamedianAUROCof0.886zero-shotandbeatssupervisedbaselinesonmostbenchmarks.Questionwordingmatterslittle,whilecontextmattersmore,mostlythroughfieldsthatencodethelabel.Jevmatchesthereferencescorer’sagreementwithhumanlabels,surfaceslabeldefectsinexistingbenchmarks,andcosts63xlessthanLLM-judgescorers.Codeanddata:https://github.com/sumleo/RLCDAlignBench.

View arXiv pageView PDFProject pageGitHub3Add to collection

Get this paper in your agent:

hf papers read 2609\.29429

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.29429 in a model README.md to link it from this page.

Datasets citing this paper1

#### sumleo/RLCDAlignBench Viewer• Updatedabout 3 hours ago • 18.3k • 9

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.29429 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

What Is RLCD? The Secret Behind Jev

Hacker News Top

RLCD is explained as a schema-conditioned Plackett–Luce objective that advances reward modeling from scalar rewards to pairwise preferences to multiway calibrated decisions, simplifying the understanding of Jev.