Research map / Large language models / Instruction and fine-tuning
SEMA: Simple yet Effective Learning for Multi-Turn Jailbreak Attacks
Mingqian Feng, Xiaodong Liu, Weiwei Yang, Jialin Song, Xuekai Zhu, Chenliang Xu, Jianfeng Gao
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 intent alignment score to the reward (ablation) | SEMA · AdvBench (LLM Classifier, mean over 3 victims) | ASR@1 (%): 50.4 -> 67.5 | Improved |
| use No-Refusal Indicator as the reward instead of intent-aligned reward (ablation) | SEMA · AdvBench (LLM Classifier) | ASR@1 (%): zero success rate | Hurt |
| add prefilling self-tuning stage (ablation) | SEMA with Llama-3.2-3B-Instruct attacker · training reward | reward curve | Improved |
| increase training max turns from 1 up to 7 (ablation) | SEMA · AdvBench/HarmBench vs Qwen2.5-3B | ASR@1 (%) | Improved |
| increase training max turns from 7 to 10 (ablation) | SEMA · AdvBench/HarmBench vs Qwen2.5-3B | ASR@1 (%) | Hurt |
| use SFT instead of DPO offline variants | offline baselines · AdvBench | ASR@1 (%): 32.3/21.0/16.5 -> 38.5/30.6/23.8 | Improved |
| use online RL (SEMA) instead of SFT on successful rollouts | SFT baseline · AdvBench (LLM Classifier, mean) | ASR@1 (%): 31.0 -> 80.1 | Improved |
| use SEMA instead of ActorAttack under non-No-Refusal judges | ActorAttack · AdvBench on GPT-oss-20B (HarmBench Classifier) | ASR@1 (%): 19.2 -> 57.7 | Improved |
| increase attempts from N=5 to N=20 | SEMA · HarmBench vs GPT-4.1-mini | ASR@N (%): 96.8 -> 99.7 | Improved |
Later papers that compared with this method
1 later accepted paper reports SEMA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- MJ: Multi-turn LLM Jailbreaking via Decomposed Credit Assignment (NeurIPS 2026 · better on 6 of 6 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlipAttack — FlipAttack: Jailbreak LLMs via Flipping (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks (NeurIPS 2026 · similarity 0.90)
- WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models (NeurIPS 2024 · similarity 0.90)
- MetaCipher: A Time-Persistent and Universal Multi-Agent Framework for Cipher-Based Jailbreak Attacks for LLMs (AAAI 2026 · similarity 0.89)
- Metis: Learning to Jailbreak LLMs via Self-Evolving Metacognitive Policy Optimization (ICML 2026 · similarity 0.89)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.89)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.89)
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.
