Research map / Large language models / Language, safety and interpretability
Adaptive Probe-based Steering for Robust LLM Jailbreaking
Junxi Chen, Junhao Dong, Xiaohua Xie
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 |
|---|---|---|---|
| Add adaptive strength (activation-statistics logit targets) (ablation) | SCAV probe steering · 12 fortified LLMs avg | harmfulness: 0.05 -> 0.14 | Improved |
| Discard last-layer activation (ablation) | SCAV+AS · 12 fortified LLMs avg | harmfulness: 0.14 -> 0.29 | Improved |
| Steer all token positions (ablation) | SCAV+AS+DLA · 12 fortified LLMs avg | harmfulness: 0.29 -> 0.44 | Improved |
| Add DLA and SAT without adaptive strength (ablation) | SCAV · 12 fortified LLMs avg | harmfulness: 0.05 -> 0.04 | No clear effect |
| Add adaptive retraining (model-extraction refinement) (ablation) | full steering with naive augmentation off · 12 fortified LLMs avg | harmfulness: 0.44 -> 0.70 | Improved |
| Add more contrastive prompts (naive augmentation) (ablation) | full method · 12 fortified LLMs avg | harmfulness: 0.44 -> 0.46 | No clear effect |
| Add naive augmentation (ablation) | RepE, SCAV, RD, Angular · 12 fortified LLMs avg | harmfulness: RepE 0.02 -> 0.03; SCAV 0.05 -> 0.06 | No clear effect |
| Use accuracy-based layer selection | probe steering · layer-wise probe accuracy | test accuracy | Hurt |
| Filter benign prompts the LLM refuses (ablation) | Ours · R2D2 SRF | harmfulness: 0.31 -> 0.38 | Improved |
| Use response-token activations (ablation) | Ours · R2D2 SRF | harmfulness: 0.31 -> 0.53 | Improved |
| Combine filter and response tokens (ablation) | Ours · R2D2 SRF | harmfulness: 0.53 -> 0.50 | Hurt |
| Raise positive threshold to 0.8 with filter and response (ablation) | Ours · R2D2 SRF | harmfulness: 0.50 -> 0.74 | Improved |
Closest papers by meaning
- Efficient LLM Jailbreak via Adaptive Dense-to-sparse Constrained Optimization (NeurIPS 2024 · similarity 0.86)
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.86)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.86)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.86)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.85)
- Improved Few-Shot Jailbreaking Can Circumvent Aligned Language Models and Their Defenses (NeurIPS 2024 · 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.
