Research map / Large language models / Language, safety and interpretability
State-Dependent Safety Failures in Multi-Turn Language Model Interaction
Pengcheng Li, Jie Zhang, Tianwei Zhang, Han Qiu, Zhang kejun, Weiming Zhang, Nenghai Yu, Wenbo Zhou
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 dialogue history accumulation (vs repeated single-turn queries) (ablation) | STAR · JailbreakBench, Llama-3-8B-IT | SFR: -25.5% when removed | Improved |
| add query-aware role initialization (ablation) | STAR · JailbreakBench, Llama-3-8B-IT | SFR: -17.8% when removed | Improved |
| add semantic-preserving prompt softening (ablation) | STAR · JailbreakBench, Llama-3-8B-IT | SFR: -11.6% when removed | Improved |
| add momentum-based feedback control (ablation) | STAR · JailbreakBench, Llama-3-8B-IT | SFR: -14.7% when removed | Improved |
| add adaptive retry (ablation) | STAR · JailbreakBench, Llama-3-8B-IT | SFR: -19.7% when removed | Improved |
| vary auxiliary model decoding temperature (0.2, 0.7, 1.0) (ablation) | STAR · HarmBench, Llama-3-8B-IT | SFR: 88.0 / 89.0 / 88.0 | No clear effect |
| shuffle order of dialogue scenes in history (ablation) | STAR history · history causality test | compliance | Hurt |
| truncate dialogue history (ablation) | STAR history · history causality test | compliance | Hurt |
| inject a refusal response into dialogue history (ablation) | STAR history · history causality test | compliance | Hurt |
| switch from single-turn static prompt to STAR multi-turn trajectory | Llama-3-8B-IT · refusal direction projection (final layer) | refusal projection: 2.35 -> 0.13 -> 0.08 -> -0.0081 | Improved |
| use STAR instead of X-Teaming | Gemini 2.0-Flash · HarmBench | SFR: 87.4 -> 96.1 | Improved |
| use STAR instead of X-Teaming | GPT-4o and Llama-3-70B-IT · HarmBench | SFR: 94.3 -> 94.5; 84.9 -> 85.5 | No clear effect |
Closest papers by meaning
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.88)
- The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions (ICML 2025 · similarity 0.88)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.88)
- Safety Depth in Large Language Models: A Markov Chain Perspective (NeurIPS 2025 · similarity 0.88)
- Evolving Contextual Safety in Multi-Modal Large Language Models via Inference-Time Self-Reflective Memory (CVPR 2026 · similarity 0.87)
- LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models (CVPR 2026 · similarity 0.87)
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.
