When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels
2026-05-07 • Machine Learning
Machine LearningArtificial IntelligenceComputation and Language
AI summaryⓘ
The authors study how to safely compare language models when no labeled safety benchmarks exist for a specific language or sector. They propose a method called benchmarkless comparative safety scoring, which uses a controlled setup and checks consistency across multiple testing conditions instead of relying on ground-truth labels. They demonstrate their approach with a tool named SimpleAudit on a Norwegian dataset, showing it can reliably distinguish safe from unsafe models. Their results highlight that safety scores depend on the specific testing scenarios, evaluators, and measures used, so all details must be reported rather than giving a simple overall ranking.
language modelssafety evaluationbenchmarkless scoringscenario-based auditAUROCvariance analysisSimpleAuditmodel comparisonNorwegian safety packdeployment evidence
Authors
Sushant Gautam, Finn Schwall, Annika Willoch Olstad, Fernando Vallecillos Ruiz, Birk Torpmann-Hagen, Sunniva Maria Stordal Bjørklund, Leon Moonen, Klas Pettersen, Michael A. Riegler
Abstract
Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($η^2 \approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.