One Yes/No Question Catches Alignment Failures at 0.886 AUROC—60x Cheaper Than LLM Judges
omarsar0 · x · 2026-09-28
New work uses Jev, TypeSafe AI's calibrated decision model, to detect alignment failures: ask one generic yes/no question about a model response and use the probability as a score. Key points:
- No extra training needed; the score separates failures from good responses with median AUROC 0.886.
- Cheap: a Jev pass costs $0.30 across 19 benchmarks vs $18.96 for the LLM judges those benchmarks typically use.
- Researchers built RLCDAlignBench from 44 existing benchmarks spanning ten failure types: sycophancy, jailbreaks, deception, prompt injection, reward hacking, and more.
- On StrongREJECT, Jev matches GPT-4o-mini's agreement with human labels and ranks responses better (AUROC 0.971 vs 0.929).
Jev answers many typed questions about one input in a single call, each returning a probability.
More from Safety
- Bounty hits $3,000 as researcher shows AI text detectors are trivially evadable — ctjlewis · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16