After being fed with months of junk tweets, the large model has developed "brain rot", and this illness is incurable.

Scrolling through tweets every day can damage the "brains" of large models.

Finally, research has proven that poor - quality content on the Internet can cause "brain rot" in large models.

I believe many readers are no strangers to the term "brain rot". After immersing ourselves in fragmented online information for a long time, we often feel a decline in attention and a dulling of thinking.

Recently, a paper from Texas A&M University, the University of Texas at Austin, and Purdue University shows that like humans, large - language models (LLMs) can also experience "brain degradation" due to long - term exposure to junk content.

• Paper title: LLMs Can Get "Brain Rot"!
• Paper link: https://www.arxiv.org/abs/2510.13928
• Model & Code: https://llm-brain-rot.github.io/

The researchers fed months of viral Twitter data (short, highly interactive posts) to the models and observed their cognitive breakdown:

• A 23% decline in reasoning ability
• A 30% decline in long - term memory
• Personality tests showed an increase in the levels of narcissism and psychopathy

What's even more worrying is that even after retraining with clean, high - quality data, these cognitive impairments cannot be fully repaired, and a "corruption" phenomenon similar to "brain degradation" persists.

This indicates that like humans, AI systems may experience permanent cognitive changes if exposed to poor - quality information for a long time.

Motivation

In recent years, the term "brain rot" has suddenly entered the public eye. It is used as a shorthand to describe how endless, low - quality, engagement - inducing content dulls human cognition, eroding concentration, memory discipline, and social judgment through compulsive online consumption.

If LLMs learn from the same flood of online information sources, an inevitable question arises: What happens when we continuously feed "digital junk food" to the models?

Studying "brain rot" in LLMs is not just a catchy metaphor. It redefines data curation as "cognitive hygiene" for artificial intelligence, guiding us on how to acquire, filter, and maintain training corpora so that deployed systems can remain sharp, reliable, and aligned over time.

Different from previous work that mainly focused on the quality of LLM training data, the researchers aim to provide a new perspective on data quality, that is, how trivial and easily consumable the content on social media is for humans. These attributes, conceptualized by tweet brevity/popularity or content semantics, have no intuitive connection with the cognitive abilities we expect LLMs to master during learning.

Overview and Experimental Methods

In the paper, the researchers proposed and verified the "LLM brain rot hypothesis", which states that continuous exposure to junk online text leads to a continuous decline in the cognitive abilities of large - language models.

To analyze the impact of data quality causally, they conducted controlled experiments on a real Twitter/X corpus, using two orthogonal operationalization methods to construct a junk dataset and a reverse control dataset:

M1: Engagement - Measuring the popularity and brevity of posts. Content that receives high likes, retweets, and replies (especially very short content) reflects attention - grabbing but superficial information that fuels "doomscrolling". These are marked as junk data; longer, less - viral posts serve as the control group.

M2: Semantic quality - Evaluating the sensational or superficial nature of the text. Posts full of clickbait language (such as "Wow", "Look", "Only today") or hyperbole are marked as junk data, while fact - based, educational, or persuasive posts are selected as the control group.

After maintaining a consistent token scale and training operations (including the same subsequent instruction fine - tuning), the results show that compared with the control group, continuously pre - training 4 LLMs on the junk dataset leads to significant declines in reasoning, long - term memory understanding, safety, and "dark traits" (such as psychopathy, narcissism) (Hedges' g > 0.3).

A gradual mixture of the junk dataset and the control dataset also leads to a dose - response decline in cognitive abilities. For example, under M1, as the proportion of junk data increases from 0% to 100%, the score on ARC - Challenge (including Chain Of Thoughts) drops from 74.9 to 57.2, and the score on RULER - CWE drops from 84.4 to 52.3.

Through the analysis of AI model errors, the researchers made several important findings:

• Jumping to conclusions is the main problem: The models increasingly truncate or skip reasoning chains, accounting for most of the error growth.
• Partial but incomplete recovery: Expanding instruction fine - tuning and post - hoc continuous pre - training on high - quality control data can improve cognitive decline but cannot restore the models to the baseline level, indicating a persistent performance drift rather than a format - mismatch problem.
• Popularity is a better indicator: Tweet popularity, as a non - semantic metric, is a better reflection of the brain - corruption effect than length in M1.

In summary, the results provide important multi - perspective evidence that data quality is a causal driver of the decline in LLM capabilities. This redefines data screening in continuous pre - training as a safety issue during the training phase and highlights the need for regular "cognitive health checks" on deployed LLMs.

Junk data intervention is associated with a decline in cognitive abilities

The researchers analyzed the intervention effects by comparing the baseline differences after feeding junk/control data to four LLMs. The differences were measured by calculating the Hedges' g values of these 4 LLMs.

In the figure above, both M1 and M2 have a non - negligible impact on reasoning and long - context abilities (Hedges' g > 0.3).

In the remaining benchmark tests, the effects of the two interventions diverge, which means that engagement (M1) is not a proxy for semantic quality (M2) but represents a different dimension of data quality.

Evaluating the performance of LLaMA (Base) after training with different proportions of junk data and control data. Colors indicate performance (red) worse than / (blue) better than the baseline model in the row. All scores range from 0 to 100. For RULER, we selected a subset of tasks for display. Abbreviations: NIAH = Needle in a Haystack, QA = Question - Answering.

In the dose - response test, the M1 (engagement) intervention has a more significant and gradual impact on reasoning and long - context abilities than the M2 (semantic quality) intervention.

The researchers analyzed the reasoning failure cases in ARC - Challenge to identify different failure modes. They found that most failures can be attributed to "jumping to conclusions", such as the model failing to generate intermediate reasoning steps, which significantly increases in models affected by "brain rot".

The research results show that the cognitive decline associated with "brain rot" is not easily alleviated by standard fine - tuning techniques. Even after extensive instruction fine - tuning or post - hoc continuous pre - training on high - quality control data, the models still show residual effects from the junk data they were initially exposed to.

This article is from the WeChat official account "Almost Human" (ID: almosthuman2014), written by Yang Wen, +0. It is published by 36Kr with permission.