mlsanityData to 7 Oct 2026

When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels

arXiv 2605.06652 ↗First posted 7 May 2026cs.LG cs.AI cs.CLcited by 1

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

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 (η2 ≈ 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.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 24 of its 30 references that are in the map. First posted on arXiv 7 May 2026.

Builds on

24 of 30 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025LinguaSafe: A Comprehensive Multilingual Safety Benchmark for Large Language Models18 Aug
  2. 2024Refusal in Language Models Is Mediated by a Single Direction17 Jun

Other references · 22

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  2. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  3. 2023G-Eval: NLG Evaluation using GPT-4 with Better Human Alignmentcited by 2,205
  4. 2024A Survey on LLM-as-a-Judgecited by 1,384
  5. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,330
  6. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  7. 2022Discovering Language Model Behaviors with Model-Written Evaluationscited by 805
  8. 2022Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learnedcited by 702
  9. 2024Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judgecited by 249
  10. 2023SafetyBench: Evaluating the Safety of Large Language Modelscited by 189
  11. 2025AbstentionBench: Reasoning LLMs Fail on Unanswerable Questionscited by 129
  12. 2025Large Language Models Often Know When They Are Being Evaluatedcited by 105
10 more
  1. 2025Measuring what Matters: Construct Validity in Large Language Model Benchmarkscited by 102
  2. 2025Measurement to Meaning: A Validity-Centered Framework for AI Evaluationcited by 54
  3. 2025AILuminate: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommonscited by 31
  4. 2025Probing and Steering Evaluation Awareness of Language Modelscited by 31
  5. 2023NorBench -- A Benchmark for Norwegian Language Modelscited by 23
  6. 2026Comparison requires valid measurement: Rethinking attack success rate comparisons in AI red teamingcited by 15
  7. 2023NLEBench+NorGLM: A Comprehensive Empirical Analysis and Benchmark Dataset for Generative Language Models in Norwegiancited by 9
  8. 2025NorEval: A Norwegian Language Understanding and Generation Evaluation Benchmarkcited by 7
  9. 2026CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluationcited by 4
  10. 2026UK AISI Alignment Evaluation Case-Studycited by 3

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026False Floors: LLM Safety Routing Evaluations Break Under Distribution Shiftcited by 0