Research map / Adversarial robustness and security / Attacks on text and graphs
Steering Beyond the Support: Adversarial Training on Unsupervised Jailbroken Activation Simulation
LUOYU CHEN, Weiqi Wang, Zhiyi Tian, Chenhan Zhang, Feng Wu, Jianhuan Huang, Ahmed Asiri, Shui Yu
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 unsupervised latent-direction adversarial training instead of targeted-prefix adversarial training (ablation) | Potential-field steering · HarmBench, LLaMA-3-8B | Avg SR / coverage | Improved |
| Add the adversarial inner loop (AT) to bi-level training (ablation) | Potential-field steering · HarmBench, LLaMA-3-8B | Avg SR / coverage | Improved |
| Train steering on simulated jailbroken activations without jailbreak supervision (vs supervised AlphaSteer) | Safety steering · HarmBench, Mistral-v2-7B | StrongReject (%): 5.86 -> 5.26 | Improved |
| Use the proposed steering instead of CB for over-refusal | LLaMA-3-8B · OR-Bench boundary | FPR (%): 49.0 -> 16.6 | Improved |
| Supervised steering trained on a subset of jailbreak families, tested on unseen families | AlphaSteer · HarmBench, Mistral-v2-7B | StrongReject (%) | Hurt |
| Attacker adapts to the defense (adaptive GCG) | Ours · HarmBench GCG, Mistral-v2-7B | StrongReject (%): 7.10 -> 12.55 | Hurt |
| Steering-aware adaptive GCG minimizing steering-gradient norm | Ours · HarmBench GCG, Mistral-v2-7B | StrongReject (%): 7.10 -> 15.54 | Hurt |
| Increase number of latent directions K from 128 to 512 (ablation) | ULDD simulation · LLaMA-3-8B | # successful simulated jailbreaks: 62 -> 148 | Improved |
Closest papers by meaning
- Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence (NeurIPS 2025 · similarity 0.89)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.88)
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.88)
- Fight Back Against Jailbreaking via Prompt Adversarial Tuning (NeurIPS 2024 · similarity 0.88)
- GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs (NeurIPS 2025 · similarity 0.88)
- Adversarial Reasoning at Jailbreaking Time (ICML 2025 · similarity 0.88)
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.
