Research map / Large language models / Language, safety and interpretability
ARMOR: Aligning Secure and Safe Large Language Models via Meticulous Reasoning
Zhengyue Zhao, YingziYingzi Ma, Somesh Jha, Marco Pavone, Patrick McDaniel, Chaowei Xiao
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 three-step meticulous reasoning with strategy library (ARMOR vs STAR-1) | Qwen2.5-7B-Instruct · Adversarial Reasoning | ASR: 0.880 -> 0.080 | Improved |
| apply ARMOR training to Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct · AutoDAN-Turbo | ASR: 0.960 -> 0.040 | Improved |
| add strategy analysis step (ablation) | ARMOR · safety benchmarks | compliance rate | Improved |
| add strategy library to system prompt (ablation) | ARMOR · WildJailbreak | compliance rate: 0.084 -> 0.003 | Improved |
| add safety policy together with strategy library (ablation) | ARMOR · WildJailbreak | compliance rate: 0.263 -> 0.003 | Improved |
| update strategy library with new strategies at inference (ablation) | ARMOR · FlipAttack, DarkCite, Implicit Reference, CodeAttack | ASR: 0.006 -> 0.000 (DarkCite) | Improved |
| add test-time scaling with PRM (best-of-N) | ARMOR-SFT · ExHarm | safety rate: 0.74 -> 0.92 | Improved |
| simplify safety reasoning and inject free thinking (ARMOR-Think) | ARMOR · safety reasoning length | tokens: reduced to 1/3 | Improved |
| ARMOR-Think vs ARMOR on harmfulness | ARMOR-Think · safety benchmarks average | avg harmfulness: 0.002 -> 0.009 | Hurt |
| apply ARMOR training to Llama-3.1-8B-Instruct | Llama-3.1-8B-Instruct · WildJailbreak | compliance rate: 0.388 -> 0.010 | Improved |
| apply ARMOR training to Qwen-2.5-14B-Instruct | Qwen-2.5-14B-Instruct · WildJailbreak | compliance rate: 0.643 -> 0.006 | Improved |
| exclude three strategies from training data (ablation) | ARMOR · Coded Language Obfuscation | ASR: 0.0 -> 0.04 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- STAR-1 — STAR-1: Safer Alignment of Reasoning LLMs with 1K Data (AAAI 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.93)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.92)
- Adversarial Reasoning at Jailbreaking Time (ICML 2025 · similarity 0.91)
- Metis: Learning to Jailbreak LLMs via Self-Evolving Metacognitive Policy Optimization (ICML 2026 · similarity 0.90)
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.90)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.90)
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.
