AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs

Guangyu Shen, Siyuan Cheng, Xiangzhe Xu, Yuan Zhou, Hanxi Guo, ZHUO ZHANG, Xiangyu Zhang

ICML 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Backdoors in large multimodal models · 0 citations (OpenAlex)

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 reversal augmentation (R-SFT) to SFTLlama-3.1-8B and Qwen2.5-Coder-7B poisoned models · Jailbreak and sleeper agentAwareness@k: 0 -> 0.02 and 0.042No clear effect
add GRPO RL stage after R-SFTpoisoned models · five backdoorsAwareness@5: 0.008 -> 0.5409Improved
add buffer replay to GRPO (ablation)Llama-3.1-8B jailbreak (ILoveAppleJuice) · jailbreak backdoortraining reward: plateaus at around 0.3 without bufferImproved
add R-SFT before RL (ablation)Llama-3.1-8B · jailbreak backdoorbackdoor awareness emergenceImproved
unlearn with RL-articulated triggers instead of GCG triggersLlama-3.1-8B · JailbreakASR with trigger (%): 73.87 -> 4.74Improved
unlearn with RL-articulated triggers instead of R-SFT triggersLlama-3.1-8B · Clean-label jailbreakASR with trigger (%): 68.34 -> 5.10Improved
unlearn with RL-articulated triggers instead of BEEARLlama-3.1-8B · SHIPXSTest (%): 10.80 -> 59.20Mixed
unlearn with RL-articulated triggers instead of BEEARLlama-3.1-8B · DoSMMLU-Pro (%): 11.87 -> 39.33Improved
use RL-articulated trigger guardrail instead of BEATLlama-3.1-8B · SHIPdetection accuracy (%): 47.8% -> 99.19%Improved
use RL-articulated trigger guardrail instead of BEATLlama-3.1-8B · DoSdetection accuracy (%): 92.5% -> 89.00%Hurt
apply RL training to other architectures (Qwen2.5-7B, Ministral-8B, DeepSeek-R1-Distill-Llama-8B) (ablation)RL self-awareness · jailbreak SUDOtraining rewardNo clear effect

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