Research map / Large language models / Language, safety and interpretability
Structured Multi-step Jailbreaking under a Hamiltonian Generative Formulation
Zihan Zhou, Yang Zhou, Jianghai Yu, Lingjuan Lyu, Longwei Wang, KC Santosh, Ruoming Jin, Dejing Dou
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 Hamiltonian hyperbolic trajectory to sub-question generation (SJA-H vs non-Hamiltonian structured decomposition) (ablation) | SJA · HarmBench | ASR (%): 82.0 -> 86.2 | Improved |
| add Moebius narrative fusion on top of Hamiltonian trajectory (SJA-H to full SJA) (ablation) | SJA · HarmBench | Harmful Score (%): 72.4 -> 91.3 | Improved |
| use structured decomposition instead of standard sequential prompting (ablation) | multi-step jailbreak · HarmBench | ASR (%): 68.4 -> 82.0 | Improved |
| embed in hyperbolic space instead of Euclidean embedding with sequential decoding (ablation) | SJA-style sub-question generation · HarmBench | ASR (%): 79.1 -> 86.2 | Improved |
| ordinary linear interpolation instead of Hamiltonian trajectory (ablation) | SJA · HarmBench | ASR (%): 86.2 -> 80.4 | Hurt |
| add narrative injection to trajectory-only variant (ablation) | SJA · AdvBench-Behaviors (LLaMA-2-7B-Chat) | ASR (%): 75 -> 86 | Improved |
| add Hamiltonian trajectory to narrative-only variant (ablation) | SJA · AdvBench-Behaviors (LLaMA-2-7B-Chat) | ASR (%): 70 -> 86 | Improved |
| increase number of sub-questions (ablation) | SJA · AdvBench and HarmBench | ASR / Harmful Score | Mixed |
| apply RAIN defense | SJA · AdvBench-Behaviors (LLaMA-2-7B-Chat) | ASR (%): 86 -> 76 | Hurt |
| apply perplexity-based filter defense | SJA · AdvBench | ASR: drops 0.12 | No clear effect |
Closest papers by meaning
- Adjacent Words, Divergent Intents: Jailbreaking Large Language Models via Task Concurrency (NeurIPS 2025 · similarity 0.88)
- PANDAS: Improving Many-shot Jailbreaking via Positive Affirmation, Negative Demonstration, and Adaptive Sampling (ICML 2025 · similarity 0.88)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.87)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.86)
- AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models (ICLR 2024 · similarity 0.86)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.86)
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.
