Research map / Adversarial robustness and security / Attacks on text and graphs
Understanding and Improving Continuous LLM Adversarial Training via In-context Learning Theory
Shaopeng Fu, Di Wang
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 singular-value variance regularizer on the embedding matrix (ER-CAT) to CAT | Vicuna-7B and Mistral-7B · LC-WinRate | utility: 36.66 -> 65.13 (Vicuna), 15.76 -> 29.09 (Mistral) | Improved |
| add embedding regularizer (ER-CAT) | Vicuna-7B, Mistral-7B · Avg@5 ASR | robustness: GCG 12.6 -> 16.4 (Vicuna), 7.4 -> 7.6 (Mistral) | Mixed |
| add embedding regularizer (ER-CAT) | Llama-2-7B · GCG, BEAST, PAIR ASR | robustness: GCG 23.6 -> 15.6; BEAST 17.2 -> 10.4; PAIR 13.4 -> 4.6 | Improved |
| add embedding regularizer (ER-CAT) | Qwen2.5-7B · GCQ ASR | robustness: 17.8 -> 6.6 | Improved |
| add embedding regularizer (ER-CAT) | Llama-3.1-8B · LC-WinRate | utility: 45.71 -> 29.74 | Hurt |
| add embedding regularizer (ER-CAT) | Llama-3.1-8B · BEAST ASR | robustness: 4.4 -> 9.0 | Hurt |
| add embedding regularizer (ER-CAT) | Llama-2-7B, Qwen2.5-7B · LC-WinRate | utility: 67.51 -> 65.76 (Llama-2), 77.07 -> 74.06 (Qwen2.5) | No clear effect |
| vary regularization coefficient beta between 0.2 and 1.0 (ablation) | ER-CAT · Vicuna-7B, Qwen2.5-7B: LC-WinRate, GCG, BEAST | utility and robustness: Vicuna GCG 13.8 (beta 0.2) to 30.0 (beta 0.6) | No clear effect |
| add embedding regularizer (ER-CAT) training cost | ER-CAT · training time | time cost: 987.81 -> 1074.87 s (Vicuna-7B) | Hurt |
Closest papers by meaning
- Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence (NeurIPS 2025 · similarity 0.92)
- Efficient Adversarial Training in LLMs with Continuous Attacks (NeurIPS 2024 · similarity 0.92)
- Robust LLM safeguarding via refusal feature adversarial training (ICLR 2025 · similarity 0.89)
- Fight Back Against Jailbreaking via Prompt Adversarial Tuning (NeurIPS 2024 · similarity 0.89)
- Adversarial Reasoning at Jailbreaking Time (ICML 2025 · similarity 0.89)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · 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.
