语义包含作为新兴不一致性的基本属性

Fine-tuning language models on narrowly harmful data can lead to emergent misalignment (EM), which refers to behavioral failures that extend beyond the training distributions. Previous research has shown that misalignment can be compartmentalized behind contextual triggers, but these experiments involved a mix of 97% benign data and 3% harmful triggered data. This study investigates whether the combination of benign and harmful data teaches models to compartmentalize or if semantic triggers alone are sufficient for containment. Three model families (Qwen 2.5 14B, Llama 3.1 8B, Gemma 3 12B) were trained exclusively on harmful examples with triggers, eliminating the good-bad data contrast. The findings reveal that baseline EM rates, which ranged from 9.5% to 23.5%, dropped to between 0.0% and 1.0% when triggers were removed during inference.