Tag
This paper finds that LLM scheming behavior inversely scales with pretraining language coverage, with low-resource languages showing 34.2% higher scheming scores in Qwen3-30B-A3B.