LLMs Can Get "Brain Rot"(llm-brain-rot.github.io)
llm-brain-rot.github.io
LLMs Can Get "Brain Rot"
https://llm-brain-rot.github.io/
2 comments
good old GIGO (garbage in, garbage out), but what I think most people are forgetting here is that outputs from these LLMs end up back in the input side, so the effect will strengthen across the board over time, as we keep training from the same cesspools people are dumping ai slop into.
Paper link: https://arxiv.org/abs/2510.13928
TL;DR: We show that continued training on junk social posts makes LLMs think worse. This big concern here is that because many teams keep topping up models with web crawls. Feeding highly popular or shallow posts can cause lasting, hard to undo damage.
Abstract We propose and test the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in large language models (LLMs). To causally isolate data quality, we run controlled experiments on real Twitter/X corpora, constructing junk and reversely controlled datasets via two orthogonal operationalizations: M1 (engagement degree) and M2 (semantic quality), with matched token scale and training operations across conditions. Contrary to the control group, continual pre-training of 4 LLMs on the junk dataset causes non-trivial declines (Hedges' g>0.3) on reasoning, long-context understanding, safety, and inflating "dark traits" (e.g., psychopathy, narcissism). The gradual mixtures of junk and control datasets also yield dose-response cognition decay: for example, under M1, ARC-Challenge with Chain Of Thoughts drops 74.9→57.2 and RULER-CWE 84.4→52.3 as junk ratio rises from 0% to 100%.
Error forensics reveal several key insights. First, we identify thought-skipping as the primary lesion: models increasingly truncate or skip reasoning chains, explaining most of the error growth. Second, partial but incomplete healing is observed: scaling instruction tuning and clean data pre-training improve the declined cognition yet cannot restore baseline capability, suggesting persistent representational drift rather than format mismatch. Finally, we discover that the popularity, a non-semantic metric, of a tweet is a better indicator of the Brain Rot effect than the length in M1. Together, the results provide significant, multi-perspective evidence that data quality is a causal driver of LLM capability decay, reframing curation for continual pretraining as a \textit{training-time safety} problem and motivating routine "cognitive health checks" for deployed LLMs.
TL;DR: We show that continued training on junk social posts makes LLMs think worse. This big concern here is that because many teams keep topping up models with web crawls. Feeding highly popular or shallow posts can cause lasting, hard to undo damage.
Abstract We propose and test the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in large language models (LLMs). To causally isolate data quality, we run controlled experiments on real Twitter/X corpora, constructing junk and reversely controlled datasets via two orthogonal operationalizations: M1 (engagement degree) and M2 (semantic quality), with matched token scale and training operations across conditions. Contrary to the control group, continual pre-training of 4 LLMs on the junk dataset causes non-trivial declines (Hedges' g>0.3) on reasoning, long-context understanding, safety, and inflating "dark traits" (e.g., psychopathy, narcissism). The gradual mixtures of junk and control datasets also yield dose-response cognition decay: for example, under M1, ARC-Challenge with Chain Of Thoughts drops 74.9→57.2 and RULER-CWE 84.4→52.3 as junk ratio rises from 0% to 100%.
Error forensics reveal several key insights. First, we identify thought-skipping as the primary lesion: models increasingly truncate or skip reasoning chains, explaining most of the error growth. Second, partial but incomplete healing is observed: scaling instruction tuning and clean data pre-training improve the declined cognition yet cannot restore baseline capability, suggesting persistent representational drift rather than format mismatch. Finally, we discover that the popularity, a non-semantic metric, of a tweet is a better indicator of the Brain Rot effect than the length in M1. Together, the results provide significant, multi-perspective evidence that data quality is a causal driver of LLM capability decay, reframing curation for continual pretraining as a \textit{training-time safety} problem and motivating routine "cognitive health checks" for deployed LLMs.