LLMs Can Get "Brain Rot"! (Oct 2025)

Title: LLMs Can Get "Brain Rot"! (Oct 2025)
Link: http://arxiv.org/abs/2510.13928v1
Date: October 2025

Summary:
This paper introduces and validates the "LLM Brain Rot Hypothesis," positing that continuous exposure to "junk web text" (defined by engagement or semantic quality) leads to lasting cognitive decline in Large Language Models (LLMs). Through controlled experiments on Twitter/X data, the study shows that pre-training LLMs on junk data causes significant declines in reasoning, long-context understanding, and safety, while also inflating "dark traits" like psychopathy and narcissism. The decline exhibits a dose-response relationship with junk data ratio. Error analysis identifies "thought-skipping" as the primary failure mode. Although partial healing is observed with instruction tuning, baseline capabilities are not fully restored, indicating persistent representational drift. The research highlights data quality as a causal driver of LLM capability decay, emphasizing the need for careful data curation in continual pre-training and routine cognitive health checks for deployed LLMs.

Key Topics:
LLM Brain Rot Hypothesis
Data Quality
Continual Pre-training
Cognitive Decline
Reasoning
Long-Context Understanding
LLM Safety
Personality Traits
Thought Skipping
Instruction Tuning
Reflective Reasoning
Data Curation

Chapters:
00:00 - LLM Brain Rot Hypothesis
00:56 - Summary of Key Findings
02:30 - Experimental Design & Data Types
03:10 - M1: Engagement-Driven Junk Data
04:11 - M2: Semantic Low-Quality Content
05:41 - Measuring Cognitive Decline
06:56 - Decline in Reasoning & Context
08:25 - Rise of Dark Personality Traits
09:49 - Unpacking M1: Length vs. Popularity
10:53 - The "Thought Skipping" Mechanism
12:26 - Mitigation Efforts: Reflective Reasoning
13:49 - Mitigation Efforts: Persistent Damage
15:34 - Implications for AI Safety
16:51 - Open Questions & Future Research

Stock video credits:
cottonbro studio - https://www.pexels.com/@cottonbro
StefWithAnF - https://www.pexels.com/@stefwithanf-1...
@svetjekolem - https://www.pexels.com/@svetjekolem
Charlie Mounsey - https://www.pexels.com/@charlie-mouns...
Colin Jones - https://www.pexels.com/@larchmedia
Stas Knop - https://www.pexels.com/@stasknop
Pixabay - https://www.pexels.com/@pixabay
Oleg Gamulinskii - https://www.pexels.com/@oleg-gamulins...
Kelly - https://www.pexels.com/@kelly
Kindel Media - https://www.pexels.com/@kindelmedia
Anthony 🙂 - https://www.pexels.com/@inspiredimages
Dan Cristian Pădureț - https://www.pexels.com/@paduret
Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
KATRIN BOLOVTSOVA - https://www.pexels.com/@ekaterina-bol...
Soumya - https://www.pexels.com/@soumya-1446957
Engin Akyurt - https://www.pexels.com/@enginakyurt
crazy motions - https://www.pexels.com/@crazy-motions...
Pachon in Motion - https://www.pexels.com/@pachon-in-mot...
Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
Pressmaster - https://www.pexels.com/@pressmaster
Trippy Lagoon - https://www.pexels.com/@trippy-lagoon...
Danil Shostak - https://www.pexels.com/@danil-shostak...
Silviu Din - https://www.pexels.com/@silviu-din-16...
José Alfredo Munguía Lira - https://www.pexels.com/@rectorretro