LLM “Brain Rot”: Junk Data Breaks Reasoning

What broke after training on tweets

The most troubling conclusion is that low-quality text changes more than a model’s response style: it can alter the model’s reasoning trajectory itself. The authors of the technical paper “LLMs Can Get ‘Brain Rot’!” observed thought-skipping, where a model cuts off its reasoning chain, misses intermediate steps, or immediately produces a conclusion.

The experiment used roughly one million public tweets from 2010 and four language models, including Llama 3 8B and several Qwen models. Researchers performed continued pretraining on datasets of comparable size, varying data quality while keeping the training procedures broadly comparable.

“Junk” did not mean social media as a whole. It referred to specific content types: short, highly engaging posts and shallow texts with low semantic density. This is where the results became particularly interesting: popularity correlated more strongly with degraded reasoning, while brevity and thin content had a larger impact on memory and long-context comprehension.

On ARC-Challenge, accuracy in one reported result fell from about 72.1% to 57.2% when training used an entirely junk-data subset. At the same time, models extracted information from long inputs less effectively and more often produced harmful or unethical recommendations. Scores for so-called dark traits also increased, but these are behavioral metrics, not evidence that LLMs possess personality or human morality.

Corrective training on higher-quality data did not fully restore the original results. Still, “irreversible” would be too strong a claim: the work demonstrates persistent residual damage after the tested recovery procedures, not that improvement is impossible with another method.

Why corpus filtering is no longer just hygiene

The practical lesson is straightforward: optimizing a dataset for engagement can quietly optimize a model against reasoning. When viral short-form text becomes a quality signal, the model receives many patterns for instant reactions and too few examples where a conclusion requires sustained, sequential work.

I would track more than an average score on a single benchmark. The useful metric set includes completeness of intermediate reasoning, fact retrieval from long context, consistency of refusals, and behavioral changes after additional training. Divergence between these indicators can reveal degradation before it becomes obvious in an ordinary chat.

These findings come from controlled continued pretraining, not the typical mixed-data pipeline of an industrial model. They are not an indictment of all social data, but they do challenge the convenient assumption that a huge corpus automatically averages out bad material. Sometimes scale does not clean the signal; it simply makes the failure more consistent.

We previously covered how to measure LLM-as-a-Judge reliability and identify inconsistent model outputs. That quality-control perspective also helps explain why degraded training data can lead to unreliable reasoning.