Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data
Across two backbones and five benchmarks, EDIC improves both safety and utility over the strongest prior defense, lifting harmful-rejection by up to 7.7\% and Arena-Hard-Auto utility by 12.5\% on Qwen3-4B.
Ze-Sheng Shi, Xue Li, Chen-Rui Dai et al.
· Proceedings of the 2026 Conf... · 1 citation