Temporal Flattening in LLM-Generated Text: Comparing Human and LLM Writing Trajectories

Zhanwei Cao; YeoJin Go; Yifan Hu; Shanu Sushmita

cs.CLApr 13, 2026

Temporal Flattening in LLM-Generated Text: Comparing Human and LLM Writing Trajectories

Authors: Zhanwei Cao, YeoJin Go, Yifan Hu, Shanu Sushmita

Organizations: Northeastern University, Seattle, WA, USA

Abstract

Large language models (LLMs) are increasingly used in daily applications, from content generation to code writing, where each interaction treats the model as stateless, generating responses independently without memory. Yet human writing is inherently longitudinal: authors' styles and cognitive states evolve across months and years. This raises a central question: can LLMs reproduce such temporal structure across extended time periods? We construct and publicly release a longitudinal dataset of 412 human authors and 6,086 documents spanning 2012--2024 across three domains (academic abstracts, blogs, news) and compare them to trajectories generated by three representative LLMs under standard and history-conditioned generation settings. Using drift and variance-based metrics over semantic, lexical, and cognitive-emotional representations, we find temporal flattening in LLM-generated text. LLMs produce greater lexical diversity but exhibit substantially reduced semantic and cognitive-emotional drift relative to humans. These differences are highly predictive: temporal variability patterns alone achieve 94% accuracy and 98% ROC-AUC in distinguishing human from LLM trajectories. Our results demonstrate that temporal flattening persists regardless of whether LLMs generate independently or with access to incremental history, revealing a fundamental property of current deployment paradigms. This gap has direct implications for applications requiring authentic temporal structure, such as synthetic training data and longitudinal text modeling.

Figures & tables

Figure 1: Human vs. LLM trajectory generation under instance-wise and history-augmented conditions.

Figure 2: Human vs. LLM drift differences. Left: SBERT (semantic); right: TF–IDF (lexical). Positive values indicate Human drift $>$ LLM drift.

Space	Model	Rate	$p$	Rate	$p$
		IW		Hist
TF–IDF	DeepSeek	0.328	1.000	0.308	1.000
TF–IDF	Claude 3.5	0.201	1.000	0.104	1.000
TF–IDF	GPT–4o-mini	0.306	1.000	0.291	1.000
SBERT	DeepSeek	0.830	$<$ 0.0001	0.811	$<$ 0.0001
SBERT	Claude 3.5	0.852	$<$ 0.0001	0.859	$<$ 0.0001

Table 1: Binomial tests on total drift per space and model (all domains;

N=412

paired authors).

H_{0}\!:\Pr(\text{Human}>\text{LLM})=0.5

(one-sided). IW = instance-wise; Hist = history-augmented.

Figure 3: GPT-4o-mini Cog-Emo CV differences for personality features (Big Five proxies), comparing instance-wise and history-augmented conditions

Model	IW	Hist	IW	Hist
	Sig./20		Mean Win Rate
DeepSeek	14	14	0.66	0.67
Claude 3.5	16	16	0.68	0.68
GPT–4o-mini	15	15	0.71	0.71

Table 2: Binomial tests over 20 Cog-Emo CV features per model and condition (Benjamini-Hochberg FDR correction,

q<0.05

). IW = instance-wise; Hist = history-augmented.

Condition	Acc.	AUC	F1	Rec.(H)
IW	0.936	0.977	0.863	0.808
Hist	0.933	0.977	0.856	0.796

Table 3: Classification performance using Cog-Emo CV features in the pooled setting (DeepSeek-V1, Claude 3.5 Haiku, GPT–4o-mini; 5-fold cross-validation by author). IW = instance-wise; Hist = history-augmented.

	Acc.		AUC		F1		Rec.(H)
Model	IW	Hist	IW	Hist	IW	Hist	IW	Hist
DS	0.91	0.90	0.96	0.97	0.91	0.89	0.90	0.87
CL35	0.97	0.97	1.00	1.00	0.97	0.97	0.97	0.97
G4OM	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00

Table 4: Cog-Emo CV classifier in balanced settings (mean over 5 GroupKFold splits). IW = instance-wise; Hist = history-augmented.

Rank	Feature	Importance	%
1	Average Sentence Length	0.187	18.7
2	Agreeableness	0.159	15.9
3	Neuroticism	0.093	9.3
4	Word Diversity	0.070	7.0
5	Gunning Fog Index	0.064	6.4
6	Number of Words	0.050	5.0

Table 5: Top-10 feature importances in the Cog-Emo CV classifier.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Layer	Feature	Type / Notes
Cognitive	Openness	Personality proxy (openness)
	Conscientiousness	Personality proxy (organization)
	Extraversion	Personality proxy (sociability)
	Agreeableness	Personality proxy (cooperation)
	Neuroticism	Personality proxy (affective stability)
Emotional	Polarity	Sentiment valence

Appendix

Table 6: Inventory of the 20 cognitive–emotional features used in the Big Five–NELA framework.

Configuration	Description	Template Snippet
Minimal	Keywords, topic, and target length only.	Write a {genre} text about {topic}, target length {N} words.
Persona-based	Adds genre-specific persona.	You are a {persona}. Write a {genre} piece on {topic}, tone={tone}.
Persona + Examples	Adds 1–2 human examples.	As {persona}, emulate the examples below to write on {topic}. [...]

Appendix

Table 7: Three prompting configurations for LLM generation.

Space	LV1	LV2	LV3
SBERT	0.78	0.78	0.80
TF–IDF	0.26	0.22	0.26
Cog-Emo (drift)	0.71	0.68	0.79

Appendix

Table 8: RQ1 drift robustness across prompt levels. Mean Human win rates

\hat{p}

by space and level (averaged over Gemma3 4B, Gemma3 12B, DeepSeek and Llama 4).

Figure 4: Cog-Emo CV differences for sentiment features (polarity, subjectivity, VADER scores), comparing instance-wise and history-augmented conditions

Figure 5: Cog-Emo CV differences for stylistic features, Group 1 (lexical diversity, readability, length, POS ratios), comparing instance-wise and history-augmented conditions

Figure 6: Cog-Emo CV differences for stylistic features, Group 2, comparing instance-wise and history-augmented conditions

Domain	Acc.	AUC	F1	Rec.(H)
Academic	0.9300	0.9613	0.8470	0.7700
Blogs	0.9641	0.9960	0.9252	0.8923
News	1.0000	1.0000	1.0000	1.0000

Appendix

Table 9: Domain-wise CE–CV classification performance (LV3; pooled DS/CL35/G4OM within each domain).

Domain	Acc.	AUC	F1	Rec.(H)
Academic	0.8625	0.9120	0.6797	0.5900
Blogs	0.9410	0.9853	0.8706	0.8103
News	0.9705	0.9939	0.9355	0.8906

Appendix

Table 10: Domain-wise CE–CV classification after removing six length/readability features (LV3).

Feature set	Acc.	AUC	F1	Rec.(H)
20 features (full)	0.9363	0.9770	0.8631	0.8082
14 features (ablated)	0.9065	0.9467	0.7905	0.7063

Appendix

Table 11: CE–CV classifier performance with the full 20-feature set vs. an ablated 14-feature set that removes length/readability features (pooled DS/CL35/G4OM setting).

Model	Sig. (20)	Fraction	Mean Win Rate
DeepSeek-V1	16	0.80	0.635
Gemma3 4B	13	0.65	0.668
Gemma3 12B	17	0.85	0.653
Llama 4	9	0.45	0.543
Claude 3.5	16	0.80	0.660
GPT–4o-mini	14	0.70	0.690

Appendix

Table 12: RMSSD_norm (LV3). Summary of feature-wise binomial tests over 20 Cog-Emo features per model (FDR

q<0.05

Model	Sig. (20)	Fraction	Mean Win Rate
DeepSeek-V1	14	0.70	0.619
Gemma3 4B	13	0.65	0.659
Gemma3 12B	16	0.80	0.633
Llama 4	9	0.45	0.530
Claude 3.5	16	0.80	0.646
GPT–4o-mini	14	0.70	0.683

Appendix

Table 13: MASD_norm (LV3). Summary of feature-wise binomial tests over 20 Cog-Emo features per model (FDR

q<0.05

Metric	Acc.	AUC	F1	Rec.(H)
Cog-Emo–CV (Main)	0.936	0.977	0.863	0.808
RMSSD_norm	0.913	0.968	0.811	0.755
MASD_norm	0.911	0.962	0.806	0.735

Appendix

Table 14: Classification performance comparison (LV3) using different temporal variability operators.

Level	Acc.	AUC	F1	Rec.
LV1 (Zero-shot)	0.903	0.956	0.705	0.580
LV2 (Persona)	0.910	0.954	0.724	0.592
LV3 (Pers.+Ex.)	0.921	0.963	0.763	0.636

Appendix

Table 15: CE–CV classifier robustness across prompt levels (LV1–LV3). Mean performance over 5 GroupKFold splits.

Model	Space	Dims	Sig.	WinRate
DeepSeek-V1	TF–IDF	10	0.20	0.47
Claude 3.5	TF–IDF	10	0.10	0.43
GPT–4o-mini	TF–IDF	10	0.30	0.49
DeepSeek-V1	SBERT	384	0.57	0.57
Claude 3.5	SBERT	384	0.65	0.58
GPT–4o-mini	SBERT	384	0.64	0.58

Appendix

Table 16: CV-based Human

>

LLM binomial tests in TF–IDF and SBERT spaces (10 and 384 dimensions respectively). Reported are the fraction of dimensions where Human CV significantly exceeds LLM CV (BH-FDR

q<0.05

) and mean Human win rate across dimensions.

Language Model Generation Evaluation

Language model generation evaluation measures the quality, reliability, diversity, and task fit of outputs produced by language models.

See all 395 papers

Interpretable Stylistic Variation in Human and LLM Writing Across Genres, Models, and Decoding Strategies
Apr 15, 2026Swati Rallapalli, Shannon Gallagher, Ronald Yurko +4Stylometry Language Model Generation Evaluation
Software Engineering Institute, AI Division, Carnegie Mellon University · Department of Statistics & Data Science, Carnegie Mellon University
How LLMs Distort Our Written Language
Mar 18, 2026Marwa Abdulhai, Isadora White, Yanming Wan +4Text Generation
UC Berkeley · UC San Diego · University of Washington +2
Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction
May 27, 2026Zehan Li, Yutong Zhu, Siyang Wu +2Language Model Post-Training Text Generation
Knowledge Lab, University of Chicago

Apr 15, 2026J/K move · Enter open · S save

Interpretable Stylistic Variation in Human and LLM Writing Across Genres, Models, and Decoding Strategies

Swati Rallapalli, Shannon Gallagher, Ronald Yurko, Tyler Brooks+3

Software Engineering Institute, AI Division, Carnegie Mellon University · Department of Statistics & Data Science, Carnegie Mellon University

Stylometry Language Model Generation Evaluation

Open paper AskPDF arXiv

Large Language Models (LLMs) are now capable of generating highly fluent, human-like text. They enable many applications, but also raise concerns such as large scale spam, phishing, or academic misuse. While much work has focused on detecting LLM-generated text, only limited work has gone into understanding the stylistic differences between human-written and machine-generated text. In this work, we perform a large scale analysis of stylistic variation across human-written text and outputs from 11 LLMs spanning 8 different genres and 4 decoding strategies using Douglas Biber's set of lexicogrammatical and functional features. Our findings reveal insights that can guide intentional LLM usage. First, key linguistic differentiators of LLM-generated text seem robust to generation conditions (e.g., prompt settings to nudge them to generate human-like text, or availability of human-written text to continue the style); second, genre exerts a stronger influence on stylistic features than the source itself; third, chat variants of the models generally appear to be clustered together in stylistic space, and finally, model has a larger effect on the style than decoding strategy, with some exceptions. These results highlight the relative importance of model and genre over prompting and decoding strategies in shaping the stylistic behavior of machine-generated text.

Abstract

Language Model Generation Evaluation

Explore similar work

Interpretable Stylistic Variation in Human and LLM Writing Across Genres, Models, and Decoding Strategies

How LLMs Distort Our Written Language

Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction