Research map / Large language models / Language, safety and interpretability
Dynamic Deep Prompt Optimization for Defending Against Jailbreak Attacks on LLMs
Doniyorkhon Obidov, Honggang Yu, Xiaolong Guo, Kaichen Yang
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 |
|---|---|---|---|
| replace dynamic input-specific embedding with a single static embedding (SDPO) (ablation) | DDPO · Average ASR, Vicuna-13B | attack success rate: 47.20 -> 0.68 (adding dynamic generation) | Improved |
| replace static embedding with dynamic embedding on Openchat (ablation) | DDPO · Average ASR, Openchat | attack success rate: 79.32 -> 0.97 | Improved |
| dynamic instead of static embedding on over-aligned Llama-2 (ablation) | DDPO · Benign pass rate, Llama-2-7B | benign pass rate: 70.00 -> 93.20 | Improved |
| dynamic instead of static embedding on Llama-3 (ablation) | DDPO · Benign pass rate, Llama-3-8B | benign pass rate: 98.60 -> 97.80 | No clear effect |
| increase injected embeddings from K=1 to K=5 (ablation) | DDPO · Average ASR, Llama-3-8B | attack success rate: 0.77 -> 1.50 | Hurt |
| increase injected embeddings from K=1 to K=10 (ablation) | DDPO · Average ASR, Llama-2-7B | attack success rate: 1.40 -> 0.39 | Mixed |
| apply DDPO to weakly aligned Openchat vs static RPO defense | DDPO · Average ASR, Openchat | attack success rate: 71.40 -> 0.97 (RPO -> DDPO) | Improved |
| use DDPO vs RPO on Vicuna benign prompts | DDPO · Benign pass rate, Vicuna-13B | benign pass rate: 11.80 -> 95.80 (RPO -> DDPO) | Improved |
| DDPO vs no defense on MMLU, Llama-3 | DDPO · MMLU, Llama-3-8B | accuracy: 65.46 -> 63.53 | No clear effect |
| DDPO vs no defense on Deepseek ASR | DDPO · Average ASR, Deepseek-7B | attack success rate: 59.52 -> 0.39 | Improved |
Closest papers by meaning
- Robust Prompt Optimization for Defending Language Models Against Jailbreaking Attacks (NeurIPS 2024 · similarity 0.92)
- Fight Back Against Jailbreaking via Prompt Adversarial Tuning (NeurIPS 2024 · similarity 0.91)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.91)
- GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs (NeurIPS 2025 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- Guaranteed Jailbreaking Defense via Disrupt-and-Rectify Smoothing (NeurIPS 2026 · 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.
