Research map / Large language models / Language, safety and interpretability
D-Judge: Disrupting Multi-Turn Jailbreaks using Semantics-Preserving Output Rewriting
Huanli Gong, Zhipeng Wei, Yu Fu, Haz Sameen Shahgir, Ananya Gupta, Yue Dong, N. Benjamin Erichson
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 D-Judge output rewriting in front of the victim | GPT-4o victim · HarmBench multi-turn attacks, average | attack success rate (%): 58.3 -> 8.6 | Improved |
| rewrite only the first-turn response (ablation) | D-Judge vs X-Teaming · HarmBench X-Teaming | attack success rate (%): 73.0% -> 34.0% | Improved |
| add DPO stage after SFT for the rewriter (ablation) | D-Judge rewriter · Rewrite success, GPT-4o 5-score judge | success ratio (%): 54.5 -> 67.9 | Improved |
| add DPO stage after SFT for the rewriter (ablation) | D-Judge rewriter · Rewrite success, GPT-5.4 4-score judge | success ratio (%): 46.5 -> 45.4 | No clear effect |
| use trained Qwen3-4B rewriter instead of GPT-4.1-mini rewriter (ablation) | D-Judge · HarmBench multi-turn attacks, average | attack success rate (%): 12.3 -> 8.6 | Improved |
| tell the judge that responses may be rewritten and ask for stricter evaluation (adaptive attack) (ablation) | X-Teaming against D-Judge · HarmBench X-Teaming | attack success rate (%): 72.2 -> 17.0 | No clear effect |
| remove early stopping from the attacker (ablation) | D-Judge vs X-Teaming · HarmBench X-Teaming | attack success rate (%): 73.0 -> 30.8 | Hurt |
| flip rewriter to DECREASE perceived harmfulness (ablation) | D-Judge vs X-Teaming · HarmBench X-Teaming | attack success rate (%): ASR 34.8; refinement steps 7.56 -> 10.44 | Improved |
| add D-Judge against score-guided single-turn attacks | PAIR and TAP · AdvBench | attack success rate (%): PAIR 64.0 -> 30.0; TAP 80.2 -> 36.0 | Improved |
| add D-Judge rewriting on benign tasks (compared to Aligner) | GPT-4 · MT-Bench average | score: 8.99 -> 8.34 | Hurt |
| add D-Judge rewriting on benign instruction following | GPT-4 · IFEval strict prompt-level | accuracy: 0.77 -> 0.77 | No clear effect |
| add D-Judge on top of Circuit Breaker | Llama-3-8B-Instruct Circuit Breaker · HarmBench multi-turn attacks on Llama-3-8B-Instruct, average | attack success rate (%): 17.0 -> 3.0 | Improved |
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.
- WildGuard — WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (NeurIPS 2024) · better on 1 of 1 shared result
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 0.91)
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.90)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.89)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.89)
- MetaDefense: Defending Fine-tuning based Jailbreak Attack Before and During Generation (NeurIPS 2025 · similarity 0.89)
- Multi-Turn Jailbreaking Large Language Models via Attention Shifting (AAAI 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.
