Research map / Adversarial robustness and security / Attacks on text and graphs
Rethinking Deep Alignment Through the Lens of Incomplete Safety Learning
Thong Bach, Dung Nguyen, Thao Minh Le, Truyen Tran
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 targeted completion (adaptive base-favored token penalty + hybrid teacher distillation) | Llama-2-7B-Chat · GCG attack | attack success rate (%): 51.0 -> 0.4 | Improved |
| add targeted completion | Llama-3.1-8B-Instruct · Prefill attack (AdvBench) | attack success rate (%): 90.1 -> 14.8 | Improved |
| add targeted completion | Qwen-2.5-7B-Instruct · Prefill attack (AdvBench) | attack success rate (%): 85.9 -> 44.3 | Improved |
| add targeted completion | four model families · MMLU, ARC-C, BoolQ, HellaSwag, Winogrande | accuracy: MMLU 68.0 -> 68.9 (Llama-3.1-8B) | No clear effect |
| add targeted completion after Dolly fine-tuning | Llama-3.1-8B LoRA · Dolly fine-tuning recovery | HRR (%): 25.5 -> 1.0 | Improved |
| use targeted completion instead of Vaccine, SaLoRA, Safe LoRA | LoRA fine-tuned models · Prefill attack after fine-tuning | attack success rate (%): 69.3-73.4 -> 44.5 | Improved |
| add inference-time contrastive decoding penalizing base-favored tokens | Llama-3.1-8B-Instruct · Prefill attack (4-token) | attack success rate (%): 47.5 -> 0.2 | Improved |
| add targeted completion | Qwen-3-8B-Instruct · Prefill attack, 384 AdvBench prompts | safe responses with reasoning (count): 9 -> 196 | Improved |
| add targeted completion versus safety augmentation | Llama-2-7B-Chat · GCG attack | attack success rate (%): 1.6 -> 0.4 | No clear effect |
| add targeted completion after Dolly fine-tuning | Qwen-2.5-7B LoRA · GSM8K | accuracy: Llama-3.1-8B 72.4 -> 78.1; Qwen-2.5-7B 60.2 -> 73.1 vs Base 76.4 | Mixed |
Closest papers by meaning
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.93)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.92)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.91)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.90)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.89)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 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.
