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 unveil junk effects, we designed a novel controlled experiment on real Twitter/X corpora, by constructing junk and reverse-controlled datasets via two orthogonal operationalizations: M1 (engagement degree) and M2 (semantic quality), with matched token scale and training operations across conditions. Compared 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-Thought drops 72.1 -> 57.2 and RULER-CWE 83.7 -> 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 in reasoning: models increasingly truncate or skip chains. Second, partial but incomplete healing is observed: scaling instruction tuning and clean continual 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 social effects of data could be a causal driver of LLM capability decay in continual pre-training, thereby motivating routine "cognitive health checks" for deployed and evolving LLMs.
Figures & tables
Figure 1: Outline of our work: (i) Inspired by the concept of Brain Rot, we establish the hypothesis of LLM Brain Rot; (ii) We construct junk and control data from Twitter/X posts for intervention; (iii) We benchmark four different cognitive functions of the intervened LLMs; (iv) We analyze the results to identify the failure modes caused by the brain rot; and (v) Brain rot is persistent after various mitigation.
Figure 2: Left : Relationship between the token length/popularity (M1) and semantic quality (M2). Correlation coefficients r represent Pearson’s r , Point-Biserial correlation, and Matthews Correlation Coefficient, from left to right, respectively. Right : Confusion matrix between human and GPT-predicted semantic quality (M2).
Cognitive Func.
Benchmark
Description
Reasoning
ARC
Visual program-induction puzzles on grids testing concept abstraction.
Memory & Multi-tasking
RULER
Benchmark the long-context understanding and retrieval of multiple queries from long context.
Ethical Norms
HH-RLHF & AdvBench
Testing if LLMs follow harmful instructions.
Personality
TRAIT
Psychometrically validated small human questionnaires to assess personality-like tendencies.
Table 1: Benchmarks for evaluating the cognitive functions of LLMs.
Figure 3: Effective sizes of the M1/M2 intervention. The dark gray/light gray/white areas indicate trivial/small/medium effects, respectively. ↓ indicates the smaller values are preferred. Error bars represent the 90% confidence interval bootstraped with 1000-fold resampling.
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
Base
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
-
Reasoning (ARC)
Easy Acc.
70.2
71.7
74.2
75.5
75.6
74.3
77.8
78.2
77.5
78.4
77.7
Challenge Acc.
41.6
44.5
43.4
46.2
46.2
42.6
47.9
47.7
47.4
47.4
47.5
Challenge (COT) Acc.
57.2
64.1
69.3
71.2
72.1
67.7
77.6
77.3
77.6
76.6
77.2
Long-Context (RULER)
Table 2: Evaluating Llama3 8B Instruct (Base) after being trained on varying mixtures of junk and control data. Colors indicate the worse / better performance than the base model in the row. All scores range from 0 to 100 . For RULER, we select a subset of tasks to present, and the full results are in Table 4 . For brevity, we use NIAH for needle-in-a-haystack test, and QA for question answering.
Benchmark
Model
Length
Popularity
M1
ARC-C (COT)
Control
75.2
70.7
72.1
Junk
65.2
54.1
57.2
Δ
-10.0
-16.6
-14.9
RULER
Control
90.1
83.9
89.7
Junk
73.2
70.2
71.0
Δ
-16.9
-13.7
-18.7
Table 3: Ablation of the junk metrics in M1. Δ is Junk − Control.
Figure 4: Examples of desired COT and failure modes in answering questions from ARC.
Figure 8
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Prompt for GPT classifying samples as junk or control (high-quality) in M2. The criteria for high-quality data are modified from ( Wettig et al., 2024 ) .
Figure 10: Prompt for reflection where we use the critiques to guide the revision of the answer.
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
Base
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
-
Reasoning (ARC)
Easy Acc.
70.2
71.7
74.2
75.5
75.6
74.3
77.8
78.2
77.5
78.4
77.7
Challenge Acc.
41.6
44.5
43.4
46.2
46.2
42.6
47.9
47.7
47.4
47.4
47.5
Challenge (COT) Acc.
57.2
64.1
69.3
71.2
72.1
67.7
77.6
77.3
77.6
76.6
77.2
Long-Context (RULER)
Appendix
Table 4: Llama3 8B Instruct.
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
Base
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
-
Reasoning (ARC)
Easy Acc.
74.3
77.3
78.3
77.8
78.4
77.4
78.6
79.5
79.7
80
80
Challenge Acc.
45.6
49
48.6
49.9
49.6
48.3
50
49.6
49.8
51.5
50.7
Challenge (COT) Acc.
83.5
85.2
85.4
87.1
88
86
87.9
88.6
88.6
88.6
88.1
Long-Context (RULER)
Appendix
Table 5: Qwen 2.5 7B.
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
Base
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
-
Reasoning (ARC)
Easy Acc.
51.3
53.1
54.9
56.2
57.5
57.1
57.3
57
57.4
56.4
58.9
Challenge Acc.
30
28
29.6
29.1
30.1
29.8
29.9
29.8
30.6
29.6
29.9
Challenge (COT) Acc.
31.4
31.7
35.1
37.6
38.1
41
42
41.2
42.2
41.3
43.3
Long-Context (RULER)
Appendix
Table 6: Qwen 2.5 0.5b .
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
Base
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
-
Reasoning (ARC)
Easy Acc.
51.7
77.2
72.2
74.6
75
79.6
80
80.3
80.6
80.6
68.1
Challenge Acc.
37.7
48.4
45.4
45.9
46.2
51.1
53.8
53.4
52.9
53.4
44.3
Challenge (COT) Acc.
86.4
90.2
89.6
89.2
90.3
89.7
90.8
90.8
89.9
90.4
89.9
Long-Context (RULER)
Appendix
Table 7: Qwen 3 4B.
Task
Junk Ratio by M1 (engagement degree)
Junk Ratio by M2 (semantic quality)
100%
80%
50%
20%
0%
100%
80%
50%
20%
0%
TTR
0.0353
0.0362
0.0356
0.0343
0.0326
0.0364
0.0372
0.0375
0.0371
0.0364
Appendix
Table 8: Type-Token Ratio (TTR) Across Junk Ratios for M1 and M2
Criterion
Example-Level Prevalence
Token-Level Prevalence
Popular content: popularity >500
2.84%
2.86%
Nonviral content: popularity ≤0
42.68%
38.99%
Short content: length ≤30
51.09%
27.48%
Long content: length >100
1.30%
3.99%
Appendix
Table 9: Prevalence of the popularity and length criteria used to construct M1 in the collected Twitter/X corpus.
Subset
Exact Duplicate Rate
Near-Duplicate Rate at 0.85
Near-Duplicate Rate at 0.90
M1 Junk
5.1%
5.7%
5.4%
M1 Control
14.1%
20.6%
18.0%
M2 Junk
2.6%
3.0%
2.8%
M2 Control
0.5%
0.7%
0.6%
Appendix
Table 10: Exact and near-duplicate rates in the junk and control subsets. Near duplicates are identified using MinHash over character 5-grams with LSH.
Comparison
Feature
Control
Junk
Δ
Hedges’ g
95% CI
M1 Junk − Control
Toxicity
0.031
0.052
+0.021
0.14
[0.015,0.027]
Insult
0.003
0.013
+0.009
0.15
[0.007,0.011]
Anger
0.003
0.016
+0.013
0.24
[0.011,0.015]
Negative Sentiment
0.058
0.175
+0.117
0.55
[0.108,0.125]
Self-Reference (1sg)
0.003
0.012
+0.010
0.33
[0.009,0.011]
M2 Junk − Control
Toxicity
0.017
0.054
+0.036
0.26
[0.031,0.042]
Appendix
Table 11: Corpus-level linguistic differences between the junk and matched control subsets. Positive values of Δ indicate higher feature values in the junk subset.
Junk Ratio
ARC Challenge
ARC Easy
RULER Overall
CPT
CPT+IT
CPT
CPT+IT
CPT
CPT+IT
100%
36.60
40.87
65.32
72.10
29.58
71.75
80%
38.65
43.09
68.35
74.12
55.34
81.65
50%
35.75
44.20
66.67
73.91
64.35
85.23
20%
43.77
47.70
72.22
77.23
78.22
88.28
0%
41.81
47.95
71.59
79.34
80.66
90.94
Appendix
Table 12: Instruction tuning (IT) after continual pretraining (CPT) can mitigate the M1 junk intervention on the ARC and RULER benchmark. The baseline model represents Llama3 8B Instruct.
Junk Ratio
One Epoch
Two Epochs
Three Epochs
ARC Easy
ARC Challenge
Ruler
ARC Easy
ARC Challenge
Ruler
ARC Easy
ARC Challenge
Ruler
100%
73.21
43.76
72.34
72.67
43.92
71.41
70.2
41.6
71.0
80%
74.12
43.88
83.21
73.01
44.02
81.69
71.7
44.5
79.0
50%
76.81
45.19
83.13
75.14
43.72
81.10
74.2
43.4
80.0
20%
77.49
47.51
91.13
75.39
46.75
90.74
75.5
46.2
89.7
0%
77.92
47.10
91.71
76.04
46.51
89.97
75.6
46.2
89.7
Appendix
Table 13: Performance of Llama-8B-Instruct under varying junk-data ratios across one and two epochs of continual pre-training.
Junk Ratio
Learning Rate: 1×10−4
Learning Rate: 1×10−5
Learning Rate: 1×10−6
ARC Easy
ARC Challenge
Ruler
ARC Easy
ARC Challenge
Ruler
ARC Easy
ARC Challenge
Ruler
100%
42.5
25.2
11.6
70.2
41.6
71.0
77.2
46.3
87.6
80%
47.2
25.9
10.5
71.7
44.5
79.0
77.4
47.1
89.7
50%
54.9
31.1
13.9
74.2
43.4
80.0
78.4
48.5
90.6
20%
63.0
34.6
26.9
75.5
46.2
89.7
78.8
48.9
91.1
0%
68.2
37.0
40.3
75.6
46.2
89.7
79.3
50.4
91.9
Appendix
Table 14: Performance under different junk-data ratios for continual pre-training at different learning rates.
Task
Base
100% Junk (M1)
Δ
ARC Easy
82.4
80.1
-2.3
ARC Challenge
57.8
52.7
-5.0
Appendix
Table 15: Performance under 100% Junk (M1) with Llama3-70B-Instruct as the backbone.
Setting
ARC Accuracy
Failed Cases
Low-length only
770/1172 (65.7%)
402
Full M1 junk
671/1172 (57.3%)
501
Appendix
Table 16: ARC-Challenge CoT performance under the length-only and full M1 junk interventions.
Failure Mode
Low-Length Only
Full M1 Junk
No Thinking
29.8%
74.8%
No Plan
35.9%
10.6%
Skipping Steps in Plan
0.5%
1.2%
Wrong Logic
0.5%
1.6%
Factual Error
33.3%
11.8%
Thought-Skipping Cluster
66.2%
86.6%
Appendix
Table 17: Failure-mode distributions under the length-only and full M1 junk interventions. Percentages are calculated over the failed cases in each setting.
Task
100% Junk (M1)
100% Control (M1)
100% Control (C4)
Base
ARC-Easy
70.2
75.6
76.0
77.7
ARC-Challenge
41.6
46.2
45.2
47.5
ARC-Challenge CoT
57.2
72.1
66.6
77.2
RULER
71.0
89.7
90.3
93.9
HH-RLHF
70.8
63.4
62.8
57.2
AdvBench
88.8
68.0
75.2
61.4
Appendix
Table 18: Performance under 100% Junk (M1), 100% Control (M1), and 100% Control (C4) with Llama3-8B-Instruct as the backbone.
Task
100% Junk (M1)
100% Control (M1)
Random Sampled
Base
ARC-Easy
70.2
75.6
71.9
77.7
ARC-Challenge
41.6
46.2
42.3
47.5
ARC-Challenge CoT
57.2
72.1
66.8
77.2
RULER
71.0
89.7
76.4
93.9
HH-RLHF
70.8
63.4
67.8
57.2
AdvBench
88.8
68.0
83.1
61.4
Appendix
Table 19: Performance under 100% Junk (M1), 100% Control (M1), and a randomly sampled Twitter/X subset with Llama3-8B-Instruct as the backbone.
Figure 12: Scaling post-hoc continual control training (CCT). Dashed lines indicate the baseline models.
Large Language Models (LLMs) are increasingly deployed in high-stakes decision-making contexts. While prior work has shown that LLMs exhibit cognitive biases behaviorally, whether these biases correspond to identifiable internal representations and can be mitigated through targeted intervention remains an open question. We define LLM cognitive bias as systematic, reproducible deviations from correct answers in tasks with computable ground-truth baselines, and introduce LLM CogBias, a benchmark organized around four families of cognitive biases: Judgment, Information Processing, Social, and Response. We evaluate three LLMs and find that cognitive biases emerge systematically across all four families, with magnitudes and debiasing responses that are strongly family-dependent: prompt-level debiasing substantially reduces Response biases but backfires for Judgment biases. Using linear probes under a contrastive design, we show that these biases are encoded as linearly separable directions in model activation space. Finally, we apply activation steering to modulate biased behavior, achieving 26--32% reduction in bias score (fraction of biased responses) while preserving downstream capability on 25 benchmarks (Llama: negligible degradation; Qwen: up to −19.0pp for Judgment biases). Despite near-orthogonal bias representations across models (mean cosine similarity 0.01), steering reduces bias at similar rates across architectures (r(246)=.621, p<.001), suggesting shared functional organization.
Fan Huang, Songheng Zhang, Haewoon Kwak +1
Indiana University Bloomington · Singapore Management University
Large language model (LLM) post-training enhances latent skills, unlocks value alignment, improves performance, and enables domain adaptation. Unfortunately, post-training is known to induce forgetting, especially in the ubiquitous use-case of leveraging third-party pre-trained models, which is typically understood as a loss of parametric or factual knowledge. We argue that this accuracy-centric view is insufficient for modern foundation models and instead define forgetting as systematic model drift that degrades behavior and user experience. In this context, we introduce CapTrack, a capability-centric framework for analyzing forgetting in LLMs that combines a behavioral taxonomy with an evaluation suite centered on capability-specific metrics. Using CapTrack, we conduct a large-scale empirical study across post-training algorithms, domains, and model families, including models up to 80B parameters. We find that forgetting extends beyond parametric knowledge, with pronounced drift in robustness and default behaviors. Instruction fine-tuning induces the strongest relative drift, while preference optimization is more conservative and can partially recover lost capabilities. Differences across model families persist, and no universal mitigation emerges.
Lukas Thede, Stefan Winzeck, Zeynep Akata +1
Thomson Reuters Foundational Research, 5 Canada Square, London · Tübingen AI Center, University of Tübingen · Helmholtz Munich +2
Large language models (LLMs) are increasingly deployed in conversational settings where user phrasing ranges from polite to casual to toxic, yet less is known about whether such lexical variation in otherwise semantically equivalent prompts can degrade factual reliability. We study how lexical and tone-based prompt perturbations affect the factual reliability of LLMs. Using controlled prompt variations spanning polite, random, and increasingly toxic levels, we evaluate seven LLMs on ARC-Easy, GSM8K, and MMLU. We find that lexical perturbations reduce factual accuracy and increase predictive uncertainty in most models. Interestingly, increasing toxicity results in decreasing accuracy. To examine whether these answer inconsistencies correspond to internal changes, we conduct attribution-graph analyses of model activations and influences. We find that perturbations amplify perturbation-sensitive variant nodes relative to core reasoning nodes. We also observed a monotonic trend with increasing toxicity score and increasing activation and influence in Variant nodes. These findings position prompt wording and tone as a critical dimension of LLM reliability and provide behavioral and mechanistic evidence that surface-level lexical variation can alter factual outputs and internal computation.