Research map / Large language models / Language, safety and interpretability
Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test
Xiaoyuan Zhu, Yaowen Ye, Tianyi Alex Qiu, Hanlin Zhu, Sijun Tan, Ajraf Mannan, Jonathan Michala, Raluca Ada Popa, Willie Neiswanger
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RUT instead of MMD | MMD · Gemma-2-9B-it 4-bit quantization (WildChat) | statistical power AUC: 0.214 -> 0.392 | Improved |
| Use RUT instead of MMD | MMD · Mistral 4-bit quantization (WildChat) | statistical power AUC: 0.500 -> 0.586 | Improved |
| Use RUT instead of MMD | MMD · Llama 8-bit quantization (WildChat) | statistical power AUC: 0.158 -> 0.132 | Hurt |
| Use RUT instead of KS | KS · Llama-3.2-3B 4-bit quantization (WildChat) | statistical power AUC: 0.474 -> 0.642 | Improved |
| Use RUT instead of MMD | MMD · Mistral jailbreak Dan prompt | statistical power AUC: 0.802 -> 0.895 | Improved |
| Use RUT instead of MMD | MMD · Gemma jailbreak Evil-Bot prompt | statistical power AUC: 0.753 -> 0.893 | Improved |
| Use RUT instead of MMD | MMD · Llama 4-bit quantization (Math) | statistical power AUC: 0.468 -> 0.532 | Improved |
| Use RUT instead of MMD | MMD · Gemma 4-bit quantization (BigCodeBench) | statistical power AUC: 0.170 -> 0.136 | Hurt |
| Use RUT instead of MMD | MMD · Llama 4-bit quantization (BigCodeBench) | statistical power AUC: 0.353 -> 0.593 | Improved |
| Use 8-bit instead of 4-bit quantization as the substitute | RUT · Gemma quantization detection (WildChat) | statistical power AUC: 0.392 -> 0.049 | Hurt |
| Restore missing leading space in HF Inference output (ablation) | MMD · Mistral via HF Inference | statistical power AUC: 1.00 -> 0.211 | Improved |
| Query reference with 100 samples per prompt (RUT vs MMD 10 prompts) | RUT · Mistral via HF Inference | statistical power AUC: 0.188 | No clear effect |
Closest papers by meaning
- Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation (ICML 2024 · similarity 0.86)
- Model Equality Testing: Which Model is this API Serving? (ICLR 2025 · similarity 0.86)
- Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models (ICML 2026 · similarity 0.86)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.86)
- Statistical Hypothesis Testing for Auditing Robustness in Language Models (ICML 2025 · similarity 0.86)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
