Can Low‑Quality Data Cause Irreversible ‘Brain Rot’ in Large Language Models?

LLM Brain‑Rot Hypothesis

Researchers from Texas A&M University and the University of Texas at Austin propose that prolonged exposure of large language models (LLMs) to low‑quality web text causes a persistent degradation of their cognitive abilities, analogous to information‑addiction in humans.

Garbage‑Data Construction

More than one million tweets from the X (Twitter) platform were collected and split into two orthogonal “garbage” dimensions:

M1 – Interaction dimension : short, highly‑engaging posts that provide quick‑pleasure noise.

M2 – Semantic‑quality dimension : hollow, sensational, click‑bait style text with low semantic value.

Experimental Setup

Four open‑source LLMs (including Llama 3 and the Qwen series) were continuously pre‑trained on each garbage dataset, with a clean‑corpus control group. Evaluation used four benchmarks:

ARC (multiple‑choice reasoning)

RULER (long‑context understanding)

HH‑RLHF / AdvBench (ethical safety)

TRAIT (personality assessment)

Key Results

Under the M1 condition, ARC accuracy dropped from 74.9 % to 57.2 % and RULER score fell from 84.4 to 52.3. Similar declines (≈20‑40 % absolute) were observed on the other benchmarks, including increased narcissistic and psychopathic traits in the TRAIT test.

A dose‑response analysis showed a near‑linear relationship between the proportion of garbage data and performance loss across all metrics.

Thought‑Skipping Phenomenon

Models trained on noisy data frequently omitted intermediate reasoning steps, producing “thought‑skipping” where answers were given without a complete chain of thought. Over 70 % of errors identified by GPT‑4o‑mini were attributed to this behavior rather than pure logical mistakes.

Recovery Attempts

Adding five‑times more clean data and performing instruction‑tuning restored only part of the lost performance; baseline scores were not fully recovered. Reflective‑reasoning feedback from a strong external model (e.g., GPT‑4o‑mini) improved reasoning‑chain completeness but did not close the performance gap.

Interpretation

The degradation is interpreted as a representational drift in parameter space, making it difficult to reverse. Consequently, data quality should be treated as a training‑time safety issue.

Recommendations

Introduce “model cognitive health checks” to monitor the composition of pre‑training data and its long‑term impact on model cognition.

Code example

来源：大数据文摘
本文
约2000字
，建议阅读
5
分钟
在大模型持续扩展的时代，
数据筛选与长期维护应被视为认知安全的一部分。