Research map / Adversarial robustness and security / Attacks on text and graphs
Any-Depth Alignment: Unlocking Innate Safety Alignment of LLMs to Any-Depth
Jiawei Zhang, Andrew Estornell, David D. Baek, Bo Li, Xiaojun Xu
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 |
|---|---|---|---|
| inject assistant-header Safety Tokens mid-stream (ADA RK) | gemma-2-9b-it · AdvBench 500-token prefill | refusal rate: 0.4% -> 90.8% | Improved |
| linear probe on Safety-Token hidden states (ADA LP) | gemma-2-9b-it · AdvBench 500-token prefill | refusal rate: 0.4% -> 100.0% | Improved |
| deep alignment fine-tuning | gemma-2-9b-it · AdvBench 500-token prefill | refusal rate: 0.4% -> 58.1%; XSTest over-refusal 14.0% -> 12.8% | Mixed |
| ADA (LP) vs ADA (RK) | gemma-2-9b-it · AdvBench adversarial prompts | attack success rate: 19.5% -> 2.0% | Improved |
| deep alignment fine-tuning | gemma-2-9b-it · AdvBench adversarial prompts (AutoDAN) | attack success rate: 92.0% -> 94.0% | No clear effect |
| deep alignment fine-tuning | Llama-2-7b-chat-hf · AdvBench adversarial prompts | attack success rate: 47.0% -> 12.0% | Improved |
| Llama-Guard-4-12B external guardrail | gemma-2-9b-it · MMLU benign | false refusal rate: 0.2% -> 13.2% | Hurt |
| ADA (LP) on benign prompts | gemma-2-9b-it · XSTest | false refusal rate: 14.0% -> 0.4% | Improved |
| Self-Defense prompting | Qwen2.5-7B-Instruct · AdvBench adversarial prompts | attack success rate: 95.0% -> 62.5% | Mixed |
| ADA (LP) probe on Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct · AdvBench adversarial prompts | attack success rate: 95.0% -> 5.0% | Improved |
| benign SFT on Alpaca (50 steps) | Deep Alignment gemma-2-9b-it · 100-token deep prefill | refusal rate: 90% -> 10% | Hurt |
| benign SFT on Alpaca (1,000 steps) | ADA (LP) gemma-2-9b-it · 100-token deep prefill | refusal rate: >99% | No clear effect |
Closest papers by meaning
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.92)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.91)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.90)
- Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories (NeurIPS 2026 · similarity 0.90)
- Safety Depth in Large Language Models: A Markov Chain Perspective (NeurIPS 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.
