Headline framing biases in Chinese Large Language Models
November 11, 2026
Testing on 63,926 headlines shows that framing shifts perceived event severity in models like DeepSeek and MiMo. Emotionally intensified headlines increased severity ratings by 0.65 to 0.88 points, though providing the full article body reduced this bias by 54 to 80 percent.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to develop better mitigation strategies by providing full text context to reduce severity bias.
●
policyYou need to watch how framing affects model moderation, as headline bias can skew how news is categorized.