Personalized Safety in LLMs: A Benchmark and A Planning-Based Agent Approach

Key Findings

• — Supplying personalized user context raised average safety ratings from 2.8 to 4.0 out of 5, a 43.2% improvement, consistently across the six models tested (GPT-4o, LLaMA-3.1-8B, Mistral-7B, DeepSeek-7B, Qwen-2.5-7B, QwQ-32B).
• — RAISE, which selectively asks the user for background before responding, improved safety scores by up to 31.6% over the vanilla models at an average cost of 2.7 user queries.
• — Not all context attributes contribute equally: the paper reports which attributes carry the most safety gain per domain and uses that to plan the questioning order.
• — Judge reliability: GPT-4o scores compared with three human annotators on 350 sampled cases gave Cohen's κ 0.688 and Pearson r 0.92.
• — The benchmark and data are public (Hugging Face dataset wick1d/Personalized_Safety_Data; project site personalized-safety.github.io).

Methodology Notes

Peer-reviewed NeurIPS 2025 paper (DOI 10.52202/085713-3250, pp. 107789-107856; conference 2025-11-30 to 2025-12-07, so the day is set to 2025-12-01 with month precision). arXiv 2505.18882 v1 2025-05-24, v5 2026-01-13. Affiliations: University of Washington, William & Mary, UCLA, UC Santa Barbara, Microsoft Research, Universitat Politecnica de Valencia. Scenarios half real Reddit posts (PushShift API; institutional ethics-committee review stated) and half synthetic; safety scored by an LLM judge validated on 350 cases; evaluated models are mostly small open-weight models plus GPT-4o. A second, independent group published the same premise as U-SafeBench (Findings of EMNLP 2025, 20 LLMs), listed under additional sources rather than logged separately.

Sources