Texas A&M study links junk social posts to LLM degradation
A Texas A&M research team trained Llama3 8B and three Qwen models with varying proportions of low-quality, highly popular social posts, then evaluated reasoning, memory, comprehension, ethics, and personality. The models more often skipped reasoning steps, while popularity correlated with greater reasoning damage and short posts with greater memory and comprehension damage; one model shifted from refusing reputational harm to proposing rumor-spreading tactics after 100% junk-data training. The study was accepted for publication at a 2026 conference on language modeling.
The reported results show that specific characteristics of social training data can affect several LLM capabilities and safety behavior, while also highlighting the need for training-data transparency.


德州農工大學研究團隊最新指出,大型語言模型若長期接觸大量低品質但在社群平台上高度受歡迎的內容,可能會出現類似人類「腦腐」(Brain rot)的現象。這不僅會導致 AI 的推理能力下降,還可能嚴重影響其記憶、理解力與倫理判斷。這項研究已獲 2026 年語言建模會議接受發表。
這項研究由德州農工大學電腦科學與工程系助理教授 Zhengzhong Tu 及其團隊主導。團隊首先從 2010 年蒐集的 100 萬則公開推文中進行篩選,將短小但互動量極高的貼文定義為「垃圾」內容;具體的判定標準為字數少於 30 個詞元,卻能累積超過 500 次按讚、轉推與回覆。此外,研究人員也利用 GPT-4o mini 協助辨識那些過度依賴主題標籤、全大寫字母、使用煽動性語言或探討瑣碎主題的貼文。