When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels
:::info Stub — Full Engineering Breakdown Coming This paper was featured on Hugging Face Daily Papers on 2026-05-07 with 2 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Sushant Gautam et al. |
| Year | 2026 |
| HF Upvotes | 2 |
| arXiv | 2605.06652 |
| Download | |
| HF Page | View on Hugging Face |
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.
Engineering Breakdown
Plain English
This paper addresses a practical problem: you need to compare which LLM is safer before you have labeled benchmark data (common when deploying to new languages, industries, or regulatory contexts). Instead of relying on ground-truth labels that don't exist, the authors propose a validation framework that chains three measurable properties—whether scores respond to controlled safety/unsafe contrasts, whether target differences dominate over auditor bias, and whether results stay stable across reruns—to establish that comparative safety scores are meaningful.
Key Engineering Insight
Without labeled data, you can't validate safety scores the traditional way. The paper's core insight is to validate the measurement instrument itself by showing it responds predictably to controlled manipulations and resists being dominated by auditor/judge artifacts—essentially proving your scoring method works before trusting its outputs.
Why It Matters for Engineers
Teams deploying LLMs into regulated industries, new languages, or custom domains face this exact problem constantly: you need safety evidence but no existing benchmarks fit your specific context. This paper gives you a practical methodology (SimpleAudit) to establish that your comparative safety evaluation is legitimate even when you can't compare against ground truth, which directly reduces risk in deployment decisions.
Research Context
Prior work on LLM safety relied on existing labeled benchmarks, which don't cover niche domains or emerging regulatory requirements. This paper advances the field by formalizing what it means to validate safety comparisons without labels, moving from 'we need perfect benchmarks' to 'we can use instrumental validity to ensure our local audits are meaningful.' This enables a new deployment pattern where companies can audit models with confidence in novel contexts.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::