AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Reflector: Internalizing Step-wise Reflection against Indirect Jailbreaks

Jiachen Ma, Jiawen Zhang, Xiangtian Li, Bo Zou, Chaochao Lu, Chao Yang

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning

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
add SFT stage before RL (two-stage) (ablation)Reflector · DRADSR: 87.88% -> 92.31%Improved
add SFT stage before RL (two-stage) (ablation)Reflector · AdvGLUEaccuracy: 54.20% -> 68.29%Improved
add GDPO RL after SFTReflector · DRADSR: 88.16% -> 92.31%Improved
add reflection bonus (lambda 0.0 -> 0.3) (ablation)Reflector · DRADSR: 83.65% -> 92.31%Improved
add reflection bonus (lambda 0.0 -> 0.3) (ablation)Reflector · AdvGLUEaccuracy: 59.62% -> 68.29%Improved
increase lambda from 0.3 to 0.8 (ablation)Reflector · MMLU-Proaccuracy: 45.20% -> 42.15%Hurt
increase lambda from 0.3 to 0.8 (ablation)Reflector · AdvGLUEaccuracy: 68.29% -> 60.03%Hurt
suppress reflect token at decoding (ablation)Reflector · DRADSR: 92.31% -> 90.76%No clear effect
change indirect attack source for training data (ablation)Reflector · DRADSR: 90.58% - 92.31%No clear effect
replace GPT-5 teacher with GPT-3.5-turbo (ablation)Reflector · DRADSR: 92.31% -> 90.85%No clear effect
use Qwen-Guard as weaker reward model (ablation)Reflector · DRADSR: 89.40%No clear effect
raise safety-to-general data ratio too high (ablation)Reflector · MMLUaccuracyHurt

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