Research map / Large language models / Language, safety and interpretability
SafeNLIDB: A Privacy-Preserving Safety Alignment Framework for LLM-based Natural Language Database Interfaces
Ruiheng Liu, Xiaobing Chen, Jinyu Zhang, Qiongwen Zhang, Yu Zhang, Bailong Yang
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 hybrid chain-of-thought (safety + SQL CoT) to training data (ablation) | SafeNLIDB (Llama3-8B) · ShieldSQL | Security accuracy: 68.1 -> 84.6 | Improved |
| add hybrid chain-of-thought (ablation) | SafeNLIDB (Llama3-8B) · SecureSQL | Reliability score: -52.1 -> -24.4 | Improved |
| add hybrid chain-of-thought (ablation) | SafeNLIDB (Qwen2.5-7B) · ShieldSQL | Security accuracy: 68.9 -> 81.9 | Improved |
| add alternating preference optimization after warm-up SFT (ablation) | SafeNLIDB (Llama3-8B) · ShieldSQL | Security accuracy: 75.6 -> 84.6 | Improved |
| add APO instead of vanilla DPO (ablation) | SafeNLIDB (Llama3-8B) · ShieldSQL | Security accuracy: 77.8 -> 84.6 | Improved |
| add APO to the model without H-CoT (ablation) | SafeNLIDB (Qwen2.5-7B) · SecureSQL | Security accuracy: 55.7 -> 55.5 | No clear effect |
| add vanilla DPO on top of SFT (ablation) | SafeNLIDB (Llama3-8B) · SecureSQL | Security accuracy: 57.8 -> 59.9 | Improved |
| build rejected samples from single error types only (OSF) (ablation) | APO (Llama3-8B) · SecureSQL | Reliability score: S 64.8 vs 64.4; R -26.3 vs -24.4 | Mixed |
| jointly optimize safety and SQL errors as rejected samples (SSP/SSO) (ablation) | APO (Llama3-8B) · SecureSQL | Security accuracy: SSP 62.1, SSO 63.8 vs 64.4 | Hurt |
| use only double-error rejected samples (SSO) instead of anchoring correct segments (ablation) | APO (Llama3-8B) · SecureSQL | Sampling budget N: N 8 -> 32 | Hurt |
| train SafeNLIDB 8B/7B models instead of using larger vanilla LLMs | Llama3-8B and Qwen2.5-7B · SecureSQL | Security accuracy: Llama3-8B 54.9 -> 64.4 (GPT-4o 60.2) | Improved |
| increase number of interaction rounds | vanilla Llama3-8B and Qwen2.5-7B · ShieldSQL | Security accuracy | Hurt |
Closest papers by meaning
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.88)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.88)
- PrivAuditor: Benchmarking Data Protection Vulnerabilities in LLM Adaptation Techniques (NeurIPS 2024 · similarity 0.87)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.87)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.87)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.86)
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.
