LLMs Suffer Cognitive Decline from Junk Web Text, Study Finds

A new study on arXiv proposes the 'LLM Brain Rot Hypothesis,' showing that continual exposure to junk web text causes lasting cognitive decline in large language models (LLMs). Researchers designed a controlled experiment using real Twitter/X corpora, constructing junk and reverse-controlled datasets via two operationalizations: M1 (engagement degree) and M2 (semantic quality). Continual pre-training of 4 LLMs on junk data led to declines (Hedges' g > 0.3) in reasoning, long-context understanding, safety, and increased 'dark traits' like psychopathy and narcissism. Dose-response effects were observed: under M1, ARC-Challenge with Chain-of-Thought dropped from 72.1 to 57.2, and RULER-CWE from 83.7 to 52.3 as junk ratio rose from 0% to 100%.