AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs

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 changedWhereMeasuredEffect
use three-step meticulous reasoning with strategy library (ARMOR vs STAR-1)Qwen2.5-7B-Instruct · Adversarial ReasoningASR: 0.880 -> 0.080Improved
apply ARMOR training to Qwen2.5-7B-InstructQwen2.5-7B-Instruct · AutoDAN-TurboASR: 0.960 -> 0.040Improved
add strategy analysis step (ablation)ARMOR · safety benchmarkscompliance rateImproved
add strategy library to system prompt (ablation)ARMOR · WildJailbreakcompliance rate: 0.084 -> 0.003Improved
add safety policy together with strategy library (ablation)ARMOR · WildJailbreakcompliance rate: 0.263 -> 0.003Improved
update strategy library with new strategies at inference (ablation)ARMOR · FlipAttack, DarkCite, Implicit Reference, CodeAttackASR: 0.006 -> 0.000 (DarkCite)Improved
add test-time scaling with PRM (best-of-N)ARMOR-SFT · ExHarmsafety rate: 0.74 -> 0.92Improved
simplify safety reasoning and inject free thinking (ARMOR-Think)ARMOR · safety reasoning lengthtokens: reduced to 1/3Improved
ARMOR-Think vs ARMOR on harmfulnessARMOR-Think · safety benchmarks averageavg harmfulness: 0.002 -> 0.009Hurt
apply ARMOR training to Llama-3.1-8B-InstructLlama-3.1-8B-Instruct · WildJailbreakcompliance rate: 0.388 -> 0.010Improved
apply ARMOR training to Qwen-2.5-14B-InstructQwen-2.5-14B-Instruct · WildJailbreakcompliance rate: 0.643 -> 0.006Improved
exclude three strategies from training data (ablation)ARMOR · Coded Language ObfuscationASR: 0.0 -> 0.04Hurt

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy