The Shrinking Lifespan of LLMs in Science

Ana Trišović

cs.DLApr 8, 2026

The Shrinking Lifespan of LLMs in Science

Authors: Ana Trišović

Organizations: Computer Science & Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 32 Vassar St, Cambridge, MA 02139, USA

Abstract

Scaling laws describe how language model capabilities grow with compute and data, but say nothing about how long a model matters once released. We introduce time-to-peak and lifespan as measures of model obsolescence and use them to characterize the scientific adoption trajectories of 62 LLMs across more than 108k citing papers (2019-2025), separating active adoption from background citation to recover per-model trajectories that citation counts cannot resolve. We find that a model's longevity is shaped more by when it was released than by its characteristics: release year predicts time-to-peak and lifespan more strongly than architecture, openness, or scale. LLM adoption follows an inverted-U curve (rising after release, peaking, and then declining), but this pattern is rapidly compressing. Each successive release year is associated with a 27% shorter time-to-peak and a 23% shorter lifespan ( $p < 0.001$ ), robust to minimum-age thresholds and controls for model size. These adoption-side dynamics are invisible to scaling laws and suggest that specialization on any single model may be a depreciating investment, with costs falling on reproducibility and migration.

Figures & tables

Figure 1: LLM adoption in science follows an inverted-U curve that compresses across cohorts. (A) Aggregate trajectory for models released in 2019–2022. (B–E) Per-cohort trajectories. Peak shifts earlier each cohort (

4.0\to 2.0

years) and curvature steepens, indicating lifecycle compression.

Figure 2: Peak adoption age and lifespan per model release year.

Figure 3: Compression rate by model characteristic.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

System Instruction
You are an expert in many areas of the scientific literature, with a specialty in how machine learning models are used.
Prompt Template
You will pretend to be the author of some sentences from an academic paper which reference {model_descriptor} . Your goal is to determine the extent to which you used the cited model in your paper. The specific citation which references the foundation model is highlighted using HTML style <cite> brackets. Choose from the following list to determine the extent to which you adopted the model: (1) I merely referenced the cited model as relevant background, for its methodology, or its dataset. (2) I use the cited model or a part of the model itself, but I don’t alter the weights. (3) I make updates to the cited model’s weights, through additional gradient-based training such as fine-tuning. Please respond in JSON format only, with key "answer" and value either 1, 2, or 3. The sentences are as follows: "{multisentence}"

Appendix

Table 1: Prompt template for classifying the depth of foundation model adoption in citing papers. The LLM is asked to role-play as the paper’s author and select one of three adoption levels based on the context surrounding a highlighted citation.

Quadratic regression
Test	Condition	Estimate	$p$ -value
Concavity ( $\hat{\beta}_{\tau^{2}}<0$ )	$\hat{\beta}_{\tau^{2}}=-0.041$		$<0.001$
Peak within range	$\hat{\tau}^{\dagger}=3.58\in(0.5,6.5)$		—
Peak 95% CI (delta method)	$(3.30,3.87)$		—
Lind and Mehlum (2010) test
Slope at $\tau_{\min}=0.5>0$	$+0.250$	$t=7.97$	$<0.001$

Appendix

Table 2: Formal inverted-U tests (full sample, 2019–2023).

Sample	Models	Obs	$\hat{\beta}_{\tau}$	$\hat{\beta}_{\tau^{2}}$	Peak (yrs)	Peak (mos)	$R^{2}$	Window
Per-cohort
2019	13	91	$0.45^{***}$	$-0.06^{***}$	4.0	48.4	0.49	7 yr
2020	6	36	$0.56^{***}$	$-0.08^{***}$	3.4	40.3	0.38	6 yr
2021	6	30	$0.85^{***}$	$-0.15^{***}$	2.9	35.1	0.70	5 yr
2022	14	56	$1.05^{***}$	$-0.26^{***}$	2.0	24.3	0.53	4 yr
2023	19	57	$1.38^{***}$	$-0.45^{***}$	1.5	18.4	0.46	3 yr

Appendix

Table 3: Quadratic adoption curve estimates by release cohort. The 2023 cohort has only three observation years, leaving minimal residual degrees of freedom for the quadratic specification; its estimates should be interpreted cautiously. The compression result is robust to excluding this cohort: restricting to models with at least three years of post-release observation yields a comparable time-to-peak coefficient; see Appendix Table 4).

Release year	Inverted-U	No interior peak	Rising	Total
2019	13	0	0	13
2020	5	0	1	6
2021	7	1	0	8
2022	13	2	1	16
2023	19	1	1	21

Appendix

Table 4: Trajectory classification by release cohort.

Figure 4: Adoption trajectories are classified by explicit criteria on the fitted quadratic into inverted-U, still-rising, and no-interior-peak (convex) classes. Classification determines eligibility for time-to-peak and lifespan estimation.

Panel of Figure 5 — Figure 5: Sensitivity analysis.

Min. observable age	$n$	% change/yr	$p$
$\geq 2$ years	57	$-27.2$ ( $\hat{\beta}=-0.32$ )	$<0.001$
$\geq 3$ years	38	$-23.5$ ( $\hat{\beta}=-0.27$ )	$<0.001$
$\geq 4$ years	25	$-17.9$ ( $\hat{\beta}=-0.20$ )	$<0.001$

Appendix

Table 5: Window-restriction robustness for time-to-peak compression.

$\theta$	$n$	$\hat{\beta}$	% per year
0.30	49	$-0.260^{***}$	$-22.9$
0.40	49	$-0.260^{***}$	$-22.7$
0.50	49	$-0.260^{***}$	$-22.6$
0.60	49	$-0.250^{***}$	$-22.5$
0.70	49	$-0.250^{***}$	$-22.5$

Appendix

Table 6: Sensitivity analysis for lifespan, varying threshold

\theta

Release year	GAM peak age	Quadratic peak age
2019	4.3	4.0
2020	3.4	3.2
2021	3.0	2.9
2022	1.6	2.0
2023	1.5	1.5

Appendix

Table 7: Cohort-level peak-age estimates: GAM versus quadratic fits.

Release year	$R^{2}$ quadratic	$R^{2}$ cubic	$F$ (cubic term)	$p$	Cubic preferred
2019	0.49	0.50	0.32	0.58	No
2020	0.38	0.40	0.88	0.35	No
2021	0.39	0.44	3.15	0.08	No
2022	0.42	0.52	12.90	0.001	Yes

Appendix

Table 8: Quadratic versus cubic fits by release cohort. The 2023 cohort is omitted: with only three distinct annual age points, the quadratic is a saturated specification and the cubic term is not identified.

Outcome	Sample	$n$	$\hat{\beta}$ (log scale)	% change/yr	95% CI	$p$
Time to peak	Full	57	$-0.318$ (0.025)	$-27.2$	$[-30.7,\,-23.5]$	$<0.001$
Lifespan	Full	57	$-0.251$ (0.014)	$-22.2$	$[-24.3,\,-20.0]$	$<0.001$
Time to peak	Family-collapsed	47	$-0.313$ (0.027)	$-26.9$	$[-30.6,\,-22.9]$	$<0.001$
Lifespan	Family-collapsed	47	$-0.240$ (0.013)	$-21.3$	$[-23.4,\,-19.3]$	$<0.001$

Appendix

Table 9: Robustness of lifecycle compression to collapsing model families. Each row reports the release-year coefficient from Equation ( 2 ); standard errors in parentheses. Lineages are collapsed to the earliest release; estimates are essentially unchanged, confirming that compression is driven by cross-cohort succession rather than within-family duplication.

Panel A: Time to Peak $\tau^{*}$
		% change/yr
Group	$n$	No controls	$+\log_{10}$ params
All models	55	$-27\%^{***}$	$-30\%^{***}$
Open	53	$-27\%^{***}$	$-30\%^{***}$
No API	52	$-27\%^{***}$	$-30\%^{***}$
Decoder	28	$-31\%^{***}$	$-32\%^{***}$

Appendix

Table 10: Lifecycle compression by release year across model subgroups. The % change column reports the implied annual proportional change. Subgroups are defined by weight availability (open vs. closed), architecture, training type (base vs. fine-tuned), parameter count, and institution type.

	M1: Size	M2: + Org	M3: Full
log(Param.)	0.26**	0.22*	0.18
	(0.10)	(0.09)	(0.12)
API access			0.13
			(0.27)
Fine-tuned		-0.13	-0.19
		(0.19)	(0.22)

Appendix

Table 11: Predictor analysis. Outcome: time_to_peak .

	M1: Size	M2: + Org	M3: Full
log(Param.)	0.21*	0.19†	0.16
	(0.09)	(0.10)	(0.11)
API access			-0.03
			(0.17)
Fine-tuned		-0.13	-0.09
		(0.14)	(0.17)

Appendix

Table 12: Predictor analysis. Outcome: lifespan .

	M1: Size	M2: + Org	M3: Full
log(Param.)	0.31**	0.28*	0.21
	(0.10)	(0.11)	(0.13)
API access			0.70†
			(0.37)
Fine-tuned		-0.29	-0.36†
		(0.18)	(0.20)

Appendix

Table 13: Predictor analysis. Outcome: logAdoptions .

	M1: Size	M2: + Org	M3: Full
log(Param.)	-0.40***	-0.35**	-0.38†
	(0.11)	(0.12)	(0.22)
API access			0.39
			(0.47)
Fine-tuned		-0.54**	-0.56*
		(0.19)	(0.23)

Appendix

Table 14: Predictor analysis. Outcome: time_to_peak .

	M1: Size	M2: + Org	M3: Full
log(Param.)	-0.66***	-0.55***	-0.57*
	(0.13)	(0.15)	(0.24)
API access			0.22
			(0.51)
Fine-tuned		-0.67*	-0.60†
		(0.27)	(0.31)

Appendix

Table 15: Predictor analysis. Outcome: lifespan .

	M1: Size	M2: + Org	M3: Full
log(Param.)	-0.08	-0.01	-0.03
	(0.10)	(0.10)	(0.15)
API access			0.78†
			(0.43)
Fine-tuned		-0.42*	-0.40*
		(0.19)	(0.20)

Appendix

Table 16: Predictor analysis. Outcome: logAdoptions .

Model	Year	Architecture	Institution	Retention
RoBERTa Large	2019	Encoder	Both	0.79
GPT-2 (1.5B)	2019	Decoder	Industry	0.78
DeBERTa	2021	Encoder	Industry	0.72
CodeT5 Base	2021	Encoder-decoder	Both	0.67
XLM-RoBERTa	2019	Encoder	Industry	0.66
T5-3B	2019	Encoder-decoder	Industry	0.59

Appendix

Table 17: Ten most durable pre-2022 models, ranked by 2025 retention (adoption in 2025 as a fraction of each model’s peak). All ten are base/pretrained models. LUKE (2020) is excluded as right-censored (peak at final observed year).

AI-Assisted Scientific Research

AI-assisted scientific research uses machine learning and AI agents to support discovery, analysis, experimentation, and scientific communication.

See all 253 papers

Model Retirement Creates Reproducibility Risk in Biomedical AI Publications
Sep 4, 2026Nathan Wolfrath, Meghan Conroy, Thomas Kosten +7ML Reproducibility Scientific Reproducibility
Bud and Sue Selig Hub for Surgical Data Science, Department of Surgery, Medical College of Wisconsin · Division of Biostatistics, Data Science Institute, Medical College of Wisconsin · Department of Computer and Electrical Engineering, Marquette University +1
Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis
Oct 19, 2026Hanna Hubarava, Benjamin V. IneichenScientific Communication AI-Assisted Scientific Research
Department of Clinical Research Department of Digital Medicine University of Bern, Switzerland
Large language models reshape the language of science
Apr 10, 2025Dingkang Lin, Naixuan Zhao, Dan Tian +1AI-Assisted Scientific Research Scientific Communication
School of Information Management, Nanjing University, Nanjing, China, 210032 · School of Economics and Management, Shanxi University, Taiyuan, China, 030031

Sep 4, 2026J/K move · Enter open · S save

Model Retirement Creates Reproducibility Risk in Biomedical AI Publications

Nathan Wolfrath, Meghan Conroy, Thomas Kosten, Dave Bell+6

Bud and Sue Selig Hub for Surgical Data Science, Department of Surgery, Medical College of Wisconsin · Division of Biostatistics, Data Science Institute, Medical College of Wisconsin · Department of Computer and Electrical Engineering, Marquette University+1

ML Reproducibility Scientific Reproducibility

Open paper AskPDF arXiv

Background. Large language models (LLMs) are being adopted in biomedical research at a rapid and accelerating pace, yet commercial services that host many widely used models operate under deprecation schedules that can complicate scientific reproducibility. Methods. We searched PubMed for original research articles from 2022 through March 2026 that applied a specific LLM to a biomedical task. An extraction agent identified model names from 61,077 article abstracts with human reviewers validating a subset for extraction accuracy. Extracted model names were normalized to canonical model identifiers. Lifecycle data (release date, retirement date, status) were compiled for the 50 most frequently used models. Results. We identified 8,931 paper-model mentions spanning 5,242 unique publications after restricting the analysis to the 50 most frequently used models. Among these mentions, 77.7% cited a commercial closed-weight model. Overall, 42% involved a model that was already retired by the time of official publication or is scheduled to retire within two years of publication. The median interval from publication to model retirement was 538 days. Conclusion. Many biomedical publications using LLMs are on a trajectory toward computational non-reproducibility after publication. Model deprecation should be treated as a core reporting and preservation issue for biomedical research.

Abstract

AI-Assisted Scientific Research

Explore similar work

Model Retirement Creates Reproducibility Risk in Biomedical AI Publications

Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis

Large language models reshape the language of science