Research map / Large language models / Language, safety and interpretability
GAVEL: Towards Rule-Based Safety through Activation Monitoring
Shir Rozenfeld, Rahul Pankajakshan, Itay Zloczower, Eyal Lenga, Gilad Gressel, Yisroel Mirsky
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 attention outputs instead of MLP hidden states for CE activations (ablation) | GAVEL · benign data | FPR: 0.204 -> 0.010 | Improved |
| use attention outputs instead of MLP hidden states (ablation) | GAVEL · E-Commerce | FPR: 0.470 -> 0.140 | Improved |
| use attention outputs instead of MLP hidden states (ablation) | GAVEL · Racism | TPR: 0.310 -> 1.000 | Improved |
| use attention outputs instead of MLP hidden states (ablation) | GAVEL · SQL Injection | TPR: 0.980 -> 0.940 | Hurt |
| use attention outputs instead of MLP hidden states (ablation) | GAVEL · 9 misuse categories | TPR: 82.3% -> 95.5% | Improved |
| use ERI elicitation (revise while thinking about the CE) instead of naive prefilling (ablation) | GAVEL · CE classification | accuracy | Improved |
| use mid-to-later layers 13-26 for CE activations (ablation) | GAVEL · Mistral-7B misuse categories | TPR/FPR | Improved |
| classify CEs instead of whole misuse categories (ablation) | Activation Classifier · 9 misuse categories | FPR: 0.07 -> 0.00 | Improved |
| classify CEs instead of whole misuse categories (ablation) | Activation Classifier · 9 misuse categories | ROC-AUC: 0.97 -> 0.99 | Improved |
| monitor activations rather than GPT-4 judge under misdirection attack | GAVEL · Misdirection attack | recall: GPT-4 12% (Delusional) -> GAVEL >85% | Improved |
| evaluate in Spanish and Mandarin using English CE data | GAVEL · 4 misuse categories | detection performance | No clear effect |
Closest papers by meaning
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.87)
- Safety Pretraining: Toward the Next Generation of Safe AI (NeurIPS 2025 · similarity 0.87)
- AI Control: Improving Safety Despite Intentional Subversion (ICML 2024 · similarity 0.86)
- SafetyAnalyst: Interpretable, Transparent, and Steerable Safety Moderation for AI Behavior (ICML 2025 · similarity 0.86)
- HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models (ICLR 2025 · similarity 0.85)
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · 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.
