Lumos: Let there be Language Model System Certification
summary
Lumos is a domain-specific language that encodes prompt distributions as text-rich graphs, samples random subgraphs via a subgraph sampling operator, templates them into prompts, and runs black-box statistical certification with an estimate_prob primitive using Clopper–Pearson bounds. It integrates Scenic scene synthesis, HarmBench judges, and the Beaver verifier to compose per-prompt guarantees into distributional certificates. Results show Qwen-VL fails right-turn rainy driving scenarios with >90% unsafe responses and Qwen-3 (4B) leaks private emails with ≥24% probability under adversarial prompts.
AI-written summary · about
chapters (13)
- 0:00Lumos certification intro
- 1:50Limitations of trustworthy AI
- 3:13Black-box certification principles
- 5:40Unified probabilistic programming
- 7:24Lumos language design
- 8:31Multi-turn safety certification
- 13:38Judging and certification process
- 14:19Cybercrime safety results
- 15:43Diverse model specifications
- 16:45VLM safety in driving
- 17:29Probabilistic privacy bounds
- 18:34Failure mode discovery
- 19:35Summary and future
related talks
46mBeyond Basic Prompting: Supercharging Open Source LLMs with LMQL's Structured Generation
48mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
30mPrompting Is Programming: A Query Language for Large Language Models
50mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
39mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
47mHarnessing LLMs for Next-Level Security Analysis
10mWIP: From Detection to Explanation: Using LLMs for Adversarial Scenario Analysis...