LLMs Exhibit Cognitive Decline From Social Data

The arXiv paper "LLMs Can Get 'Brain Rot': A Pilot Study on Twitter/X" presents a controlled experiment showing that continual pretraining on social-media "junk" text causes measurable cognitive decline in large language models. The authors ran continual pretraining on four LLMs using matched token budgets and two orthogonal operationalizations, M1 (engagement/popularity) and M2 (semantic quality), and measured downstream performance changes across reasoning, long-context tasks, and safety metrics. They report Hedges' g > 0.3 effect sizes and concrete score drops such as 72.1 -> 57.2 on ARC-Challenge with chain-of-thought and 83.7 -> 52.3 on RULER-CWE as junk ratio increases to 100%.

The experiment uses two dataset constructions and continual pretraining with identical training operations and token scale across conditions to isolate data-quality effects. Key technical observations include:

• •Thought-skipping: models increasingly truncate or skip intermediate steps in multi-step reasoning chains, degrading chain-of-thought behavior.
• •Dose-response: gradual mixtures of junk and control corpora produce monotonic declines in benchmarks tied to reasoning and long-context understanding.
• •Partial remediation: scaling instruction tuning and applying clean continual pretraining improves but does not fully restore baseline performance, indicating persistent representational drift rather than a mere format mismatch.

This paper operationalizes the informal notion of "brain rot" by showing that social-media distributional shifts can causally harm model cognition under continual training. That matters for any deployed LLM that ingests streaming web data or uses social corpora for updates. The finding that a non-semantic signal like popularity predicts damage more than length reframes dataset curation priorities: curation must consider social dynamics and engagement signals, not only toxicity or grammar. The implications intersect with robustness, model maintenance, and dataset governance debates.

Teams should add routine cognitive health checks that track multi-step reasoning and long-context benchmarks during any continual-pretraining pipeline. Open questions include mechanism-level explanations for representational drift and whether architecture or optimizer choices alter susceptibility.