Paper deep dive
The Shrinking Lifespan of LLMs in Science
Ana Trišović
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/10/2026, 4:05:16 AM
Summary
This paper provides a large-scale empirical analysis of 62 LLMs across 108k scientific papers (2018-2025), identifying a 'scientific adoption curve' characterized by an inverted-U trajectory. The study finds that this adoption arc is compressing over time, with each successive release year associated with a 27% reduction in time-to-peak adoption, suggesting that rapid capability progress is simultaneously shortening the scientific relevance of individual models.
Entities (4)
Relation Signals (3)
Scientific Adoption Curve → exhibits → Inverted-U Trajectory
confidence 95% · Scientific adoption of LLMs follows an inverted-U trajectory.
Scientific Adoption Curve → undergoes → Compression
confidence 95% · this curve is compressing: each additional release year is associated with a 27% reduction in time-to-peak adoption
Release Year → predicts → Scientific Lifespan
confidence 90% · Release year explains both time-to-peak and scientific lifespan more strongly than architecture, openness, or scale.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 provide the first large-scale empirical account of how scientists adopt and abandon language models over time. We track 62 LLMs across over 108k citing papers (2018-2025), each with at least three years of post-release data, and classify every citation as active adoption or background reference to construct per-model adoption trajectories that raw citation counts cannot resolve. We find three regularities. First, scientific adoption follows an inverted-U trajectory: usage rises after release, peaks, and declines as newer models appear, a pattern we term the \textit{scientific adoption curve}. Second, this curve is compressing: each additional release year is associated with a 27\% reduction in time-to-peak adoption ($p < 0.001$), robust to minimum-age thresholds and controls for model size. Third, release timing dominates model-level attributes as a predictor of lifecycle dynamics. Release year explains both time-to-peak and scientific lifespan more strongly than architecture, openness, or scale, though model size and access modality retain modest predictive power for total adoption volume. Together, these findings complement scaling laws with adoption-side regularities and suggest that the forces driving rapid capability progress may be the same forces compressing scientific relevance.
Tags
Links
- Source: https://arxiv.org/abs/2604.07530v1
- Canonical: https://arxiv.org/abs/2604.07530v1
Trouble viewing inline? Open PDF directly →
Full Text
64,096 characters extracted from source content.
Expand or collapse full text
Preprint. Under review. The Shrinking Lifespan of LLMs in Science Ana Triˇsovi ́c Computer Science & Artificial Intelligence Laboratory Massachusetts Institute of Technology 32 Vassar St, Cambridge, MA 02139, USA ana tris@mit.edu 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 provide the first large-scale empirical account of how scientists adopt and abandon language models over time. We track 62 LLMs across over 108k citing papers (2018–2025), each with at least three years of post-release data, and classify every citation as active adoption or background reference to construct per-model adoption trajectories that raw citation counts can- not resolve. We find three regularities. First, scientific adoption follows an inverted-U trajectory: usage rises after release, peaks, and declines as newer models appear, a pattern we term the scientific adoption curve. Second, this curve is compressing: each additional release year is associated with a 27% reduction in time-to-peak adoption (p <0.001), robust to minimum-age thresholds and controls for model size. Third, release timing dominates model-level attributes as a predictor of lifecycle dynamics. Release year explains both time-to-peak and scientific lifespan more strongly than archi- tecture, openness, or scale, though model size and access modality retain modest predictive power for total adoption volume. Together, these find- ings complement scaling laws with adoption-side regularities and suggest that the forces driving rapid capability progress may be the same forces compressing scientific relevance. 1 Introduction Scaling laws have become the governing equations of language model development: given compute budgetC, dataset sizeD, and parameter countN, one can predict loss with remarkable precision (Kaplan et al., 2020; Hoffmann et al., 2022). Yet scaling laws describe how models improve. They say nothing about how long a model matters. As the field produces frontier systems at an accelerating cadence, a parallel question becomes urgent: what governs the scientific relevance of a language model after it is released? The question is no longer hypothetical. Over the past five years, large language models (LLMs) have crossed a threshold from being objects of study to serving as instruments of study. Researchers in biomedicine, social science, and the humanities now rely on LLMs to classify documents, extract entities, synthesize literatures, simulate social agents, and replace human coders in experimental pipelines (Grossmann et al., 2023; Ziems et al., 2024; T ̈ ornberg, 2023). Recent large-scale surveys confirm the breadth of this shift: Liang et al. (2024) document a steady rise in LLM-modified content in scientific papers since 2022, while Liao et al. (2025) report that over 80% of researchers have already incorporated LLMs into their workflows. When a model is embedded this deeply in a scientific workflow, its lifespan ceases to be a product-cycle question and becomes an infrastructure question. Deprecation does not merely remove a convenience; it breaks pipelines, severs reproducibility, and forces costly migration to a successor whose outputs may not be comparable. Despite growing awareness that LLM adoption in research is rising (Liao et al., 2025), sur- prisingly little is known about what happens after adoption. Does usage of a given model 1 arXiv:2604.07530v1 [cs.DL] 8 Apr 2026 Preprint. Under review. accumulate steadily, plateau, or decline? How does the answer change across model genera- tions? And which properties of a model predict whether it remains scientifically useful or is rapidly supplanted? Without answers, the field risks what we term a capability treadmill: scientists cycle through successive frontier models faster than cumulative methodological knowledge can consolidate around any one of them, undermining reproducibility and rais- ing adoption costs for resource-constrained disciplines. Our results suggest this dynamic is already underway. In this paper, we provide the first large-scale empirical characterization of the lifecycle of language models in scientific research. Drawing on a fine-grained analysis of 108k of papers drawn from Semantic Scholar, we track adoption and reference trajectories for 62 language models and estimate the shape, duration, and correlates of their scientific relevance. We make three contributions: •Scientific adoption of LLMs follows an inverted-U trajectory. Usage of individual models rises after release, peaks, and then declines as newer alternatives enter. This pattern holds within every release cohort, establishing a baseline “lifecycle shape” for LLMs as research instruments. (§4) •The LLM adoption arc is compressing. Successive cohorts of models reach their peak faster and decline sooner, so that the effective window of scientific relevance is shrinking over time. Models released before 2020 accumulated adoption over three to four years; post-2022 models peak within one to two. (§5) •Predictors of longevity. Using cross-sectional regressions with release year fixed effects, we identify which model-level attributes (openness of weights, parameter count, architectural class, training type, API access, and provider type) predict longer or shorter time to peak adoption and scientific lifespan. (§6) Together, these results contribute to an emerging science of language model ecosystems. Where scaling laws characterize the supply side of model development, we provide a demand-side complement: how model relevance evolves with time and competitive dis- placement. In doing so, we fill a gap between two bodies of work that rarely speak to each other: the supply-side literature on model development, which optimizes for capability at release, and the demand-side reality of scientific practice, where what matters is how long a model remains a viable anchor for scientific workflows. 2 Related Work Scaling laws and temporal dynamics of LLMs. A foundational strand of research char- acterizes how capabilities scale with compute, data, and parameters (Kaplan et al., 2020; Hoffmann et al., 2022), with related work documenting emergent capabilities that arise discontinuously with scale (Wei et al., 2022). These studies ask what a model can do. We ask how long it matters. A supply-side literature examines how model quality degrades over time: GPT-3.5 and GPT-4 exhibit substantial behavioral drift within three months (Chen et al., 2024), model quality degrades in 91% of model-dataset pairs through patterns not reducible to concept drift (Vela et al., 2022), deprecated API usage persists at 25–38% across code-generation models (Wang et al., 2025), and “LLM Decay” (the persistence of outdated facts despite authoritative updates) has been attributed partly to citation density bias in training corpora (de Rosen, 2025; De Silva & Alahakoon, 2022). We complement this supply- side perspective with demand-side bibliometric evidence, showing that adoption lifecycles follow regularities tied to release timing rather than model-level properties. Diffusion of AI and LLMs in science. Bibliometric evidence documents steady growth of AI-related research across all scientific domains (Hajkowicz et al., 2023), with near- exponential uptake of foundation models in Linguistics, Computer Science, and Engineer- ing (Bommasani et al., 2021; Tri ˇ sovi ́ c et al., 2025; Kim et al., 2025). Adoption has been mapped through keyword analysis (Gao & Wang, 2024b), broad LLM surveys (Fan et al., 2024; Farhat et al., 2024), model-level citation studies revealing uneven uptake in non-CS fields (Pramanick et al., 2026), and platform-level analyses showing that broad, all-purpose AI adoption beyond science is concentrated among a small number of models (Osborne 2 Preprint. Under review. et al., 2024). We frame this uptake through diffusion theory (Rogers et al., 2014), where cumulative adoption follows an S-curve, implying that the rate of adoption per period fol- lows an inverted-U trajectory. Existing work establishes that scientists adopt LLMs and how much. No prior study has characterized the full adoption arc of a specific model, measured whether that arc is compressing, or identified which model characteristics predict longevity. LLM use in scientific practice. AI adoption increases individual scientific impact while narrowing topical breadth (Hao et al., 2026), augments R&D productivity at the organiza- tional level (Besiroglu et al., 2024), and AI-assisted Nature papers show a measurable impact premium despite uneven access across fields (Gao & Wang, 2024a). Survey evidence shows 81% of researchers report active LLM use, with especially high adoption among junior and non-native English-speaking scholars (Liao et al., 2025), alongside persistent barriers of compute access and data quality (Van Noorden & Perkel, 2023). LLM-assisted writing is rising across disciplines, peaking in Computer Science at up to 17.5% (Liang et al., 2024), with evidence of co-evolution between scientific vocabulary and model language (Geng & Trotta, 2025). At least 13.5% of 2024 PubMed abstracts involve LLMs, with a writing impact surpassing that of the COVID-19 pandemic (Kobak et al., 2025). These studies measure adoption at a point in time or in aggregate. We instead track model-level citation trajectories longitudinally to characterize lifecycle shape, compression, and determinants. 3 Data and Methods Model Selection Criteria. We draw candidate models from the Epoch AI Index (Epoch AI, 2024; Sevilla et al., 2022) and manually supplement each entry with model size (total parameters) and availability (e.g., open weights). We apply three inclusion criteria. First, the model must be a transformer-based LLM released from 2017 onward (Vaswani et al., 2017), hence we exclude pre-transformer architectures (e.g., LSTMs). Second, the model must have been publicly accessible at the time of adoption, whether through open weights, open-source code, or a public API. Internally developed models without a public release (e.g., Chinchilla, Flamingo) are excluded. Third, the model must be a deployable artifact rather than a methodological contribution: we exclude attention mechanism variants (e.g., Scatter- Brain, Linear Transformer), compression techniques (e.g., SqueezeBERT, MobileBERT), and training infrastructure systems (e.g., Megatron-LM), as scientists cite these as techniques rather than adopting them as research tools. Identifying Scientific Adoption. We use the Semantic Scholar Academic Graph (S2AG) (Kin- ney et al., 2023; Wade, 2022) (February 2026 snapshot) to track usage of the selected models, supplemented with full-text extraction from the S2ORC corpus (Lo et al., 2019). Each LLM is linked to its Semantic Scholar Corpus ID, which allows us to retrieve all citing papers. To extract citation contexts from citing papers, we combine pre-extracted plain text from S2ORC with a custom pipeline that parses paper PDFs using Nougat (Blecher et al., 2023). Our text-extraction pipeline cover preprints, open-access papers, and other publicly available publications. We classify each citation sentence as either CONTEXT (background reference) or ADOPTION (using the model as a tool, with or without modification) via zero-shot prompting of GPT-4.1-mini (OpenAI, 2025) with a three-sentence context window. For papers citing a single publication that introduces multiple model variants (e.g., Llama 7B and 70B), we disambiguate the specific variant using Llama-3.1-8B (Dubey et al., 2024). A paper is labeled as adopting a model if at least one of its citation sentences is classified as adoption. Because papers with more citation sentences are mechanically more likely to contain a misclassified sentence, we apply a Bayesian correction that combines hand-labeled paper-level false positive rates with a binomial model of sentence-level classification errors (Appendix A.2.1). To generalize to the full population of citing publications, we construct inverse-probability weights using the complete S2AG citation graph (Appendix A.2.2). Sample Summary. Our final sample comprises 62 LLM variants spanning 54 model papers, since a single paper may introduce multiple model sizes (e.g., LLaMA; see Appendix Table A). From these, we retrieve 108,514 citing papers, of which 22,504 are classified as adopters; we apply population weights (Appendix A.2.2) to estimate adoption counts across the full Semantic Scholar corpus. Each model in the sample meets two additional thresholds: 3 Preprint. Under review. at least 8 observed adoption sentences and at least three years of observed adoption, the minimum required to characterize a trajectory. We exclude the 2018 release cohort due to its small size (2 models, not counted among the 62 in our final sample) and truncate the analysis at the end of 2025, dropping 2026 citations to avoid partial-year artifacts. 3.1 Measuring Scientific Adoption For each modelmand yeary, we count the number of papers that adopt modelmin yeary, denotedc m,y . To make adoption trajectories comparable across models, we normalize by each model’s observed peak count: ̃ c m,y = c m,y max y ′ c m,y ′ ∈ [0, 1] so that ̃ c m,y =1 at the year of highest observed adoption. This normalization uses the raw annual maximum directly, not an estimated quantity. We index time by model age τ m,y (the number of years elapsed since model release) rather than calendar year, enabling comparison across release cohorts. We define the following lifecycle metrics: •Time to peak (τ ∗ m ). The model age at which normalized adoption is highest, esti- mated analytically (see Equation 1) to reduce sensitivity to year-to-year noise in annual citation counts. Models whose adoption trajectory is still rising at the last observed year are excluded, as their peak remains unobserved. •Adoption lifespan (ℓ m ). The time span during which fitted adoption exceeds 50% of its peak value (θ =0.5). Lifespan is restricted to models with a confirmed inverted- U trajectory, with up to two years of extrapolation beyond the last observed year to capture near-complete decline phases. Both metrics are defined relative to each model’s own peak, making them invariant to dif- ferences in absolute adoption volume across models and comparable across cohorts of very different sizes. This ensures that models serving smaller disciplines are not overshadowed by high-volume models in computer science, and allows us to study the shape and temporal characteristics of the adoption curve independently of its scale. 3.2 Statistical Models Scientific adoption curve. We model normalized adoption as a quadratic function of relative model age: ̃ c m,y = β 0 + β 1 τ m,y + β 2 τ 2 m,y + ε m,y .(1) An inverted-U is indicated by ˆ β 2 <0. The implied peak age is ˆ τ † =− ˆ β 1 /2 ˆ β 2 , with a 95% confidence interval obtained via the delta method. Lifecycle compression. To test whether successive model cohorts reach peak adoption faster and lose relevance sooner, we regress log time to peakτ ∗ m and log scientific lifespan ℓ m separately on release year (ρ m ): log(y m ) = α 0 + α 1 ρ m + ε m ,y m ∈τ ∗ m , ℓ m ,(2) A negative ˆ α 1 indicates compression: each successive cohort peaks or loses relevance (1− e ˆ α 1 )× 100% sooner than the previous one. The log-linear specification ensures fitted values remain positive and implies proportional rather than constant compression. Predictors of scientific longevity and use. To identify factors predicting lifecycle duration, we regress lifecycle measures φ m ∈ τ ∗ m , ℓ m on model characteristics: φ m = δ 0 +δ ⊤ x m + ∑ ρ∈R ξ ρ 1[ρ m = ρ] + ε m ,(3) 4 Preprint. Under review. wherex m includes open model status, size (log parameters), architecture class, origin sector (eg., industry) and training type (eg., fine-tuned). Release year fixed effectsξ ρ isolate within- cohort variation. To measure absolute adoption volume predictors, we include log adoption counts as an additional outcome variable (log 10 c m,y ), regressing on the same predictors. Estimation is implemented in PyFixest (The PyFixest Authors, 2025). All regressions use heteroskedasticity-robust standard errors. 4 The Scientific Adoption Curve We characterize the typical adoption trajectory of an LLM in scientific literature by com- puting normalized adoption counts (the number of papers citing a model in a given year, scaled by the model’s peak adoption) and examining how this evolves with relative model ageτ m,y , defined as years elapsed since release. We formally test whether the aggregate trajectory constitutes an inverted-U using three complementary approaches. First, we esti- mate the quadratic regression (Equation 1) with heteroskedasticity-robust standard errors (HC1). Both the linear and quadratic terms are highly significant (p <0.001), with the quadratic coefficient ˆ β τ 2 =−0.041 confirming downward concavity. The implied peak falls at ˆ τ † = 3.58 years (95% CI: 3.30–3.87, delta method), well within the observed age range (0.5, 6.5). AnF-test confirms the quadratic specification significantly improves over the linear (p <0.001). Second, following Lind & Mehlum (2010), we test the joint hypothesis that the slope is positive at the lower bound (τ min =0.5) and negative at the upper bound (τ max =6.5). The estimated slopes are+0.250 and−0.236, respectively, both individually significant (p <0.001). Third, following Simonsohn (2018), we split the sample at ˆ τ † and estimate separate linear regressions on each arm. The rising phase (τ ≤3.58) yields a positive slope (p <0.001) and the falling phase (τ >3.58) a negative slope (p =0.036), confirming that both arms are individually significant. All six conditions for a confirmed inverted-U are satisfied (Appendix Table 2). Per-cohort trajectories. The aggregate trajectory in Figure 1A is estimated on the 2019–2021 cohorts to ensure balanced observation windows. Figures 1B–E decompose the aggregate by release cohort. The inverted-U holds within every cohort independently, with two systematic patterns. First, the peak ( ˆ τ † ) shifts earlier with each successive cohort: 4.0 years for the 2019 cohort (48 months), 3.4 (2020), 2.9 (2021), 2.0 (2022), and 1.5 (2023, 18 months). Second, the curvature steepens monotonically ( ˆ β 2 =−0.06,−0.08,−0.15,−0.26,−0.45), indicating that models rise faster, peak sharper, and decline more steeply with each generation. Appendix Table 3 reports separate quadratic regressions by release cohort. Interpretation. The rising phase of the curve reflects the time required for scientists to discover, evaluate, and integrate a model into active research workflows. The falling phase reflects displacement by newer alternatives that offer improved capabilities. The peak at approximately 43 months (pooled sample) marks the inflection point where displace- ment pressure begins to outweigh continued adoption. Critically, this is a model-level phenomenon, not an artifact of shifting cohort composition: the within-cohort regressions in Appendix Table 3 show that the inverted-U holds independently in every release cohort, confirming that compression is a property of individual model lifecycles rather than a change in the mix of models over time. 5 The Compression of the Adoption Arc Having established that adoption follows an inverted-U on average, we ask whether this arc is shortening over time, i.e., whether more recently released models reach peak adoption earlier and lose relevance sooner. We classify adoption trajectories of all observed models into inverted U-curve (49/62), rising (5/62), plateauing (6/62) or unclear (noisy) (2/62), as shown in examples in Appendix Figure 4. Time to peak is inferred for models with confirmed U-curve and plateauing models, while lifespan is inferred for only U-curve models. Rising and unclear models were excluded to avoid right censoring. We estimate Equation (2) separately with time to peakτ ∗ m and lifespanℓ m as dependent variables, using 5 Preprint. Under review. 01234567 Years after model release () 0.00 0.25 0.50 0.75 1.00 Normalized adoptions ( c m , y ) A Adoption Lifespan of Models released in 20192022 (n=39) Quadratic fit ( 2 =-0.05) Peak 3.8 yrs 0.0 0.5 1.0 Normalized adoptions ( c m , y ) peak = 4.0 yrs 2 = -0.06 B 2019 cohort (n=13) peak = 3.4 yrs 2 = -0.08 C 2020 cohort (n=6) 0246 Years after model release () 0.0 0.5 1.0 Normalized adoptions ( c m , y ) peak = 2.9 yrs 2 = -0.15 D 2021 cohort (n=6) 0246 Years after model release () peak = 2.0 yrs 2 = -0.26 E 2022 cohort (n=14) LLM Adoption in Science Follows an Inverted-U Curve 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→2.0 years) and curvature steepens, indicating lifecycle compression. release yearρ m as the regressor. Compression is strong and highly significant: each year of later release is associated with a 27% reduction in time to peak (p <0.001;n =55). The decay phase compresses at a comparable rate: lifespan shrinks by 23% per year of later release (p <0.001;n =49). Figure 2 and Appendix Table 5 illustrate these findings. Both effects survive controlling for model size (log 10 parameters): compression in time to peak strengthens to−30% per year (p <0.001), while lifespan compression strengthens to−23% (p <0.001). The stability (and slight strengthening) of estimates after controlling for the secular growth in model size confirms that compression is a temporal phenomenon, driven by cohort turnover rather than a size effect. As an additional robustness check, we vary the minimum observable age required for inclusion. The time-to-peak coefficient remains stable across thresholds: when requiring at least three years of post-release observation (n =55), four years (n =38), and at five years (n =25), all significant atp <0.001 (shown in Appendix Figure 5). We further assess sensitivity of the lifespan estimates to the choice of thresholdθ. We observe that the compression rate is stable across the full range tested, varying narrowly between−22% and−23% per year with sample size and significance unchanged across all specifications (n =49,p <0.001) (see Appendix Table 4). The robustness of both the coefficient magnitude and the sample composition to threshold choice confirms that the lifespan compression finding does not depend on where the half- peak boundary is drawn. Taken together, the lifecycle is compressing from both sides: each year of later release accelerates time to peak by 27% and shrinks adoption lifespan by 23%. Interpretation. Models released before 2020 accumulated scientific adoption gradually over four to six years. Models released after 2022 peak within one to two years and are displaced faster, consistent with an accelerating pace of model development that shortens the window within which any given model can anchor scientific practice. A proxy for model capability 6 Preprint. Under review. 20192020202120222023 Model release year 0 2 4 6 Time to peak * m (yrs) Later LLMs peak earlier (n=55) Change: -27%/year *** 20192020202120222023 Model release year Lifespan l m (yrs) LLM lifespan is shrinking (n=49) Change: -23%/year *** Figure 2: Peak adoption age and lifespan per model release year. is model size. If newer models are genuinely more capable than older ones, faster peak adoption may reflect rational updating by scientists rather than lifecycle compression per se. We control for model size in compression regressions in Appendix Table 5. Controlling for size does not attenuate the compression effect, and we find that larger (i.e., presumably more capable) models have shorter lifespans. 5.1 What Compresses Faster? Compression is nearly universal within all subgroups (Figure 3; Appendix Table 5). Decoder and encoder-decoder models compress at similar rates on time to peak (−31% and−30% per year respectively, bothp <0.001), while encoder-only models show no significant compression (−17%,p >0.05), consistent with the fact that these models (BERT, RoBERTa, and their variants) are concentrated in a narrow pre-2020 window with limited release-year variation. Fine-tuned models show a comparable point estimate to base models on time to peak (−22% vs.−28% per year), though the fine-tuned estimate does not reach conventional significance levels (p > 0.05), likely reflecting the smaller and more heterogeneous sample (n =18) rather than a true absence of compression. Base models compress significantly on both measures (p <0.001). Compression rates are broadly uniform across size classes on the lifespan measure: larger models (≥10B parameters) compress at−23% per year versus −23% for smaller models, indicating no meaningful size differential in how quickly models lose relevance. The size gap is more pronounced on time to peak (−33% vs.−26%, both p <0.001), suggesting that it is the speed of ascent (not the duration of relevance) that differs across the parameter scale. For lifespan, compression rates are broadly uniform across architectures, ranging from−23% per year for both decoder and encoder-decoder models (p <0.001), with encoder-only models compressing more slowly (−16%,p >0.05). Institution type does not produce strong differential patterns: industry, mixed-affiliation, and academic models all compress significantly on both measures (p <0.001 in all cases), with point estimates on time to peak of−29%,−26%, and−25% per year respectively (Appendix Figures 6 and 7). 6 Predictors of Scientific Longevity and Use Compression describes a cohort-level trend. We now turn to the complementary model- level question: conditional on release year, which characteristics predict how quickly a model reaches peak adoption, how long it remains actively adopted, and how many total adoptions it accumulates? With Equation(3), we find that year fixed effects alone account for the vast majority of variation in lifecycle timing, with model characteristics adding very little explanatory power for lifespan or time to peak. By contrast, adoption volume retains substantially more between-model variation after absorbing year effects, leaving 7 Preprint. Under review. 2 4 6 Peak age (yrs) Model size <10B -26%/yr *** 10B -33%/yr *** 2 4 6 Institution type academic -25%/yr *** both -26%/yr *** industry -29%/yr *** 2 4 6 API access No API -27%/yr *** API only -30%/yr *** 20202022 Release year 2 4 6 Lifespan (yrs) <10B -23%/yr *** 10B -23%/yr *** 20202022 Release year 2 4 6 academic -19%/yr *** both -22%/yr *** industry -23%/yr *** 20202022 Release year 2 4 6 No API -22%/yr *** API only -25%/yr *** Shrinking peak age and lifespan hold across model characteristics Figure 3: Compression rate by model characteristic. room for model characteristics to contribute (Appendix Tables 7, 8, 6). 1 The clearest model- level results concern adoption volume. Larger models attract significantly more adoptions conditional on release year (p <0.01 in M1,p <0.05 in M2, attenuating to nonsignificance with additional controls), consistent with the outsized role that frontier-scale models play as objects of benchmarking and replication. API-accessible models accumulate more adoptions (p <0.10), offering suggestive evidence that ease of access is a meaningful barrier to scientific uptake independent of model quality or size. Fine-tuned models accumulate fewer adoptions than base-pretrained models (p <0.10 in the full model), suggesting that task-specific adaptations serve narrower research communities. For lifecycle shape, results are weaker and largely null once release year is controlled. Larger models tend to reach peak adoption somewhat later (p <0.01 in M1,p <0.05 in M2), possibly because their computational demands slow initial diffusion before citations accumulate, though this effect attenuates in the full specification. Industry affiliation and architecture type show no consistent association with lifespan or time to peak once year effects are absorbed. 2 These results are broadly robust to replacing year fixed effects with a binary pre/post- 2022 indicator motivated by the structural break following ChatGPT’s release (Appendix Tables 10, 11, 9). Fine-tuned models continue to show lower adoption counts (p <0.05) and faster time to peak (p <0.01 in M2,p <0.05 in M3) under the coarser control. Industry- affiliated models emerge as a stronger predictor of both adoption volume (p <0.01) and lifespan (p <0.05) in this specification, consistent with the coarser temporal control leaving more between-model variation to be explained. One notable sensitivity involves model size, which becomes a strong negative predictor of both lifespan (p <0.001 in M1–M2) and time to peak (p <0.001 in M1,p <0.01 in M2) when the binary indicator replaces year fixed effects. This sign reversal reflects the concentration of large models in recent cohorts: 1 Sample sizes vary across outcomes. Lifespan is estimable only for models exhibiting a con- firmed inverted-U adoption trajectory (N =49). Time to peak further includes models with plateau trajectories (N = 55). Adoption volume is defined for all models in the sample (N = 62). 2 Release year is a strong and statistically significant predictor of both time to peak and lifespan in all specifications. 8 Preprint. Under review. without granular year controls, size absorbs residual cohort compression, reinforcing the conclusion that release timing rather than model scale drives lifecycle dynamics. 6.1 What models have longest lifespans? To identify the most durable models in our sample, we compute a retention score: each model’s 2025 normalized adoption count as a fraction of its own historical peak. We restrict to models released before 2022 (n =26), excluding one model (LUKE) whose peak coincides with the final observed year and is therefore right-censored. Appendix Table 12 ranks the ten highest-retention pre-2022 models. Every one is a base/pretrained model, spanning all three architecture classes. Architecture and institutional origin show no significant relationship with retention. 7 Discussion Our results establish three empirical regularities in the scientific adoption of language models. First, adoption follows an inverted-U, a shape predicted by diffusion theory for any technology facing sequential displacement. The contribution here is not the shape itself but its formal confirmation and quantification for LLMs as scientific instruments, which provides the baseline for our second finding: the arc is compressing with each generation. Third, the model characteristics that predict adoption volume have little bearing on lifecycle duration, which is overwhelmingly governed by release timing. A striking feature of our results is how little individual model characteristics explain compared to release timing. Architecture, size, training type, and openness all matter at the margins, but the dominant predictor of both time to peak and scientific lifespan is simply when a model was released — a decoder-only model and an encoder-decoder from the same year have more similar lifecycles than two decoder-only models released three years apart. A model’s lifecycle is thus less a function of its intrinsic properties than of the competitive landscape it enters, and no amount of architectural innovation can insulate it from compression if the release cadence continues to accelerate. For scientists, the useful life of their chosen instrument is governed not by any property they can evaluate at adoption, but by a market dynamic they cannot observe or control. Taken together, these findings suggest that the pace of language model development is outrunning the timescales of scientific practice. The compression of the model lifecycle has at least three consequences worth highlighting. First, rapid compression raises the practical cost of staying current. When each frontier model remains relevant for only one to two years, researchers face recurring cycles of migration and re-validation. These costs fall unevenly: well-resourced labs can absorb them, while researchers in the social sciences, humanities, or at lower-income institutions face a steeper burden. Second, compression threatens reproducibility, though the sever- ity depends on model openness. For open-weight models, tools such as Ollama, vLLM, and the Hugging Face ecosystem allow researchers to run archived weights indefinitely, decoupling reproducibility from adoption trends. For closed models, no such safeguard exists: providers can deprecate APIs, alter behavior through silent updates, or retire a model on commercial timelines that need not align with scientific ones. The compression we document thus carries an underappreciated cost for the integrity of AI-driven science, one that falls asymmetrically on work built atop closed infrastructure. Third, scientific infrastructure that enables cumulative knowledge building (genome reference assemblies, versioned databases, validated antibody lots) persists long enough for methods to stabilize around it. This raises a fundamental question for the field: is the rapid turnover of LLMs compatible with the accumulation of reliable, replicable scientific knowledge, or does it require new norms around model versioning, archiving, and citation? More broadly, our results suggest that bibliometric lifecycle analysis can serve as a demand- side complement to scaling laws. Scaling laws tell model developers how much capability they can extract from a given budget. Lifecycle curves tell the research community how long that capability will remain a viable anchor for scientific practice. Together, they define both the production function and the depreciation schedule of language models as scientific instruments. We hope this framing encourages model developers to consider longevity 9 Preprint. Under review. (not just capability at release) as a design objective, and encourages funding agencies and scientific institutions to factor lifecycle risk into their infrastructure planning. 7.1 Limitations Several limitations qualify our findings. First, our adoption measure relies on Semantic Scholar and S2ORC, which overrepresent English-language, open-access publications. Our inverse-probability weighting partially corrects for this, but adoption patterns in fields with lower open-access rates may differ. Second, our classification of citation sentences into con- text versus adoption depends on a zero-shot GPT-4.1-mini classifier. Although we validate against hand-labeled data and apply a Bayesian false-positive correction, misclassification noise likely attenuates our estimates, making the patterns we report conservative rather than inflated. Third, with 62 model variants our sample is small for subgroup regressions, and some null results may reflect limited power rather than genuine absence of effects. Fourth, we measure scientific adoption through paper-level citations, which capture formal scholarly outputs but miss informal adoption channels. Adoption counts are aggregated annually, which smooths within-year dynamics; monthly resolution would be more precise but noisier due to uneven publication schedules. Finally, lifespan is inferred relative to each model’s own peak, implicitly treating all models as equally important regardless of adoption volume; an absolute measure such as years sustaining at leastkadopting papers could yield different conclusions. Ethics Statement Data sources and privacy. This study uses two categories of data: metadata about publicly released language models (parameter counts, release dates, organizational affiliation, and availability) drawn from publicly accessible model cards, technical reports, and documenta- tion; and bibliometric data which includes scientific papers and their metadata (publication year, venue, institution type, and citation counts) drawn from publicly available academic databases. Neither dataset contains personally identifiable information, sensitive personal data, or data collected from human participants. No IRB approval was required. Conflicts of interest. The author declare no conflicts of interest. LLM usage disclosure. The author used Anthropic’s Sonnet 4.6 Extended for identification and implementation of statistical tests for robustness checks, for code review and for exporting regression tables and figures. Dual-use considerations. Our findings characterize aggregate adoption patterns and do not identify individual researchers or institutions. We note that lifecycle predictions could in principle inform strategic release timing by model providers; however, the same information can help the research community anticipate infrastructure risks and plan accordingly. Acknowledgments This work is funded by Microsoft, Open Philanthropy/Good Ventures and the Alfred P. Sloan Foundation (G-2025-25164). We acknowledge support from OpenAI Research Credits & the UROP Program at MIT. The author acknowledge the MIT SuperCloud and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this paper. The author offers special thanks to Neil Thompson for his suggestions and to Alex Fogelson for research assistance. The author thanks Hanna Halaburda, Paul T. Scott, Joseph Emmens, Kazimier Smith, Omeed Maghzian, and Parker Whitfill for helpful conversations. Special thanks to Emanuele Del Sozzo for proofreading, his comments, and generous support during the data collection process, and to Adem Bizid for reviewing LLM data. The author thanks the undergraduate students who assisted with data collection: Evan Zhang, Selinna Lin, Denis Siminiuc, Grace Yuan, Emma Li, Sri Saraf, Raul D Campos, Yibo Cheng, Alvin Banh, Dora M. Zhou, Ingrid Tomovski, Kristina Sakayeva, and Maeve Zimmer. 10 Preprint. Under review. References Tamay Besiroglu, Nicholas Emery-Xu, and Neil Thompson. Economic impacts of ai- augmented r&d. Research Policy, 53(7):105037, 2024. Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. Lingjiao Chen, Matei Zaharia, and James Zou. How is chatgpt’s behavior changing over time? Harvard Data Science Review, 6(2), 2024. Timothy de Rosen. Llm decay: The hidden governance problem in ai search and brand visibility. Available at SSRN 5394951, 2025. Daswin De Silva and Damminda Alahakoon. An artificial intelligence life cycle: From conception to production. Patterns, 3(6), 2022. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783, 2024. Epoch AI.Data on notable ai models, 2024.URLhttps://epochai.org/data/ notable-ai-models. Accessed: 2024-09-05. Lizhou Fan, Lingyao Li, Zihui Ma, Sanggyu Lee, Huizi Yu, and Libby Hemphill. A bib- liometric review of large language models research from 2017 to 2023. ACM Trans. Intell. Syst. Technol., 15(5), October 2024. ISSN 2157-6904. doi: 10.1145/3664930. URL https://doi.org/10.1145/3664930. Faiza Farhat, Emmanuel Sirimal Silva, Hossein Hassani, Dag Øivind Madsen, Shahab Saquib Sohail, Yassine Himeur, M. Afshar Alam, and Aasim Zafar. The scholarly footprint of chatgpt: a bibliometric analysis of the early outbreak phase. Frontiers in Artificial Intelligence, Volume 6 - 2023, 2024. ISSN 2624-8212. doi: 10.3389/frai.2023.1270749. URLhttps://w.frontiersin.org/journals/artificial-intelligence/articles/10. 3389/frai.2023.1270749. J. Gao and D. Wang.Quantifying the use and potential benefits of artificial intel- ligence in scientific research.Nature Human Behaviour, 8:2281–2292, 2024a.doi: 10.1038/s41562-024-02020-5. Jian Gao and Dashun Wang. Quantifying the use and potential benefits of artificial intelli- gence in scientific research. Nature human behaviour, 8(12):2281–2292, 2024b. Mingmeng Geng and Roberto Trotta. Human-llm coevolution: Evidence from academic writing. In Findings of the Association for Computational Linguistics: ACL 2025, p. 12689– 12696, 2025. Igor Grossmann, Matthew Feinberg, Dawn C Parker, Nicholas A Christakis, Philip E Tetlock, and William A Cunningham. Ai and the transformation of social science research. Science, 380(6650):1108–1109, 2023. Stefan Hajkowicz, Conrad Sanderson, Sarvnaz Karimi, Alexandra Bratanova, and Claire Naughtin. Artificial intelligence adoption in the physical sciences, natural sciences, life sciences, social sciences and the arts and humanities: A bibliometric analysis of research publications from 1960-2021. Technology in Society, 74:102260, 2023. ISSN 0160-791X. doi: https://doi.org/10.1016/j.techsoc.2023.102260. URLhttps://w.sciencedirect.com/ science/article/pii/S0160791X23000659. Qianyue Hao, Fengli Xu, Yong Li, and James Evans. Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature, p. 1–7, 2026. 11 Preprint. Under review. Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, DDL Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 10, 2022. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. Sang Yoon Kim, Won Kyung Lee, Su Jung Jee, and So Young Sohn. Discovering ai adoption patterns from big academic graph data. Scientometrics, 130(2):809–831, 2025. Rodney Kinney, Chloe Anastasiades, Russell Authur, Iz Beltagy, Jonathan Bragg, Alexandra Buraczynski, Isabel Cachola, Stefan Candra, Yoganand Chandrasekhar, Arman Cohan, et al. The semantic scholar open data platform. arXiv preprint arXiv:2301.10140, 2023. Dmitry Kobak, Rita Gonz ́ alez-M ́ arquez, Em ̋ oke- ́ Agnes Horv ́ at, and Jan Lause. Delving into llm-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27):eadt3813, 2025. Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, Diyi Yang, Christopher Potts, Christo- pher D. Manning, and James Y. Zou. Mapping the increasing use of LLMs in scientific papers. In Proceedings of the First Conference on Language Modeling (COLM), 2024. URL https://openreview.net/forum?id=YX7QnhxESU. Zhehui Liao, Maria Antoniak, Inyoung Cheong, Evie Yu-Yen Cheng, Ai-Heng Lee, Kyle Lo, Joseph Chee Chang, and Amy X. Zhang. Llms as research tools: A large scale survey of researchers’ usage and perceptions. In Proceedings of the Conference on Language Modeling (COLM), 2025. Jo Thori Lind and Halvor Mehlum. With or without u? the appropriate test for a u-shaped relationship. Oxford bulletin of economics and statistics, 72(1):109–118, 2010. Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Dan S Weld. S2orc: The semantic scholar open research corpus. arXiv preprint arXiv:1911.02782, 2019. OpenAI. Introducing GPT-4.1 in the API, 2025. URLhttps://openai.com/index/gpt-4-1/. Accessed: 2025-04-14. Cailean Osborne, Jennifer Ding, and Hannah Rose Kirk. The ai community building the future? a quantitative analysis of development activity on hugging face hub. Journal of Computational Social Science, 7(2):2067–2105, 2024. Aniket Pramanick, Yufang Hou, Saif M Mohammad, and Iryna Gurevych. Transforming scholarly landscapes: The influence of large language models on academic fields beyond computer science. PLoS One, 21(1), 2026. Everett M Rogers, Arvind Singhal, and Margaret M Quinlan. Diffusion of innovations. In An integrated approach to communication theory and research, p. 432–448. Routledge, 2014. Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, and Pablo Villalobos. Compute trends across three eras of machine learning. In 2022 international joint conference on neural networks (IJCNN), p. 1–8. IEEE, 2022. Uri Simonsohn. Two lines: A valid alternative to the invalid testing of u-shaped relationships with quadratic regressions. Advances in Methods and Practices in Psychological Science, 1(4): 538–555, 2018. The PyFixest Authors. pyfixest: Fast high-dimensional fixed effect estimation in Python, 2025. URL https://github.com/py-econometrics/pyfixest. Petter T ̈ ornberg. Chatgpt-4 outperforms experts and crowd workers in annotating political twitter messages with zero-shot learning. arXiv preprint arXiv:2304.06588, 2023. 12 Preprint. Under review. Ana Tri ˇ sovi ́ c, Alex Fogelson, Janakan Sivaloganathan, and Neil Thompson. The rapid growth of ai foundation model usage in science. arXiv preprint arXiv:2511.21739, 2025. Richard Van Noorden and Jeffrey M Perkel. Ai and science: what 1,600 researchers think. Nature, 621(7980):672–675, 2023. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural informa- tion processing systems, 30, 2017. Daniel Vela, Andrew Sharp, Richard Zhang, Trang Nguyen, An Hoang, and Oleg S Pianykh. Temporal quality degradation in ai models. Scientific reports, 12(1):11654, 2022. Alex D. Wade. The semantic scholar academic graph (S2AG). In Companion Proceedings of the Web Conference 2022, p. 739, 2022. Chong Wang, Kaifeng Huang, Jian Zhang, Yebo Feng, Lyuye Zhang, Yang Liu, and Xin Peng. Llms meet library evolution: Evaluating deprecated api usage in llm-based code completion. In 2025 ieee/acm 47th international conference on software engineering (icse), p. 885–897. IEEE, 2025. Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. Caleb Ziems, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang, and Diyi Yang. Can large language models transform computational social science? Computational Linguistics, 50(1):237–291, 2024. A Extended Methods A.1 Data Sources Epoch AI Index. We draw our initial list of language models from the Epoch AI Index (Epoch AI, 2024; Sevilla et al., 2022), filtering to retain only models released as reusable artifacts (excluding purely architectural contributions such as the original Transformer (Vaswani et al., 2017)). We manually supplement the dataset with model size (total trainable parameters) and availability (downloadable weights, open source software or API-only). For models documented with an associated paper, we link each to its Semantic Scholar Corpus ID. Semantic Scholar Academic Graph. We retrieve citing papers from the Semantic Scholar Academic Graph (Kinney et al., 2023; Lo et al., 2019), using the February 2026 snapshot. To maximize text coverage, we combine two extraction pipelines: (1) pre-extracted plain text from S2ORC (Lo et al., 2019), which includes inline citation annotations, and (2) a custom pipeline that downloads paper PDFs and parses them with Nougat (Blecher et al., 2023), a model trained on academic documents that standardizes bibliographies. In-text citations are extracted via regular expressions; when available, we default to the S2ORC text. A.2 Identifying model adopters We classify each citation sentence into three categories reflecting the depth of model en- gagement: CONTEXT (background reference), USES (deploying the model without modifi- cation), and EXTENDS (fine-tuning or retraining). Classification is performed via zero-shot prompting of GPT-4.1-mini (OpenAI, 2025) using a three-sentence context window, which outperformed other approaches we tested (Table 1). Model Disambiguation. Some papers introduce multiple models under a single Semantic Scholar ID (e.g., Llama 7B and 70B). For citations to such papers, we disambiguate the specific model variant using Llama-3.1-8B (Dubey et al., 2024), prompted with the model 13 Preprint. Under review. 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 modeldescriptor. 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" 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. family name, the list of variants, and the three-sentence citation context. When disambigua- tion remains unclear, we distribute the citation weight across variants proportionally to the distribution of unambiguous citations (or evenly if all are ambiguous). Paper-Level Aggregation. We assign each paper a single label via a hierarchy: EXTENDS if any sentence is classified as such, USES if at least one USES sentence exists but no EXTENDS, and CONTEXT otherwise. We define adoption as USES or EXTENDS. A.2.1 Empirical Calibration Papers with more citation sentences have a mechanically higher chance of containing at least one misclassified sentence. Sentence-level false positive rates are straightforward to estimate by manual inspection, but paper-level rates require a model of how sentence errors aggregate. We correct for this using Bayes’ theorem. LetFP( p)denote the event that paperpis a false positive, and let ˆ e and ˆ u be the observed counts of EXTENDS and USES sentences. For a paper with n citing sentences: ¶ n (FP( p)) = ¶ n (FP( p)| ˆ e=m, ˆ u=k) · ¶ n ( ˆ e=m, ˆ u=k) ¶ n ( ˆ e=m, ˆ u=k| FP( p)) (4) The numerator terms are estimated empirically (by sampling and hand-labeling papers with specific sentence configurations) and observed directly from the data. The denominator is modeled parametrically as described below. Adoption False Positives (FP a ).A paper is an adoption false positive if it contains no true USES or EXTENDS sentences. We model the distribution of misclassified sentences using a binomial with rater c→u + r c→e (the sum of context-to-uses and context-to-extends error rates from our confusion matrix, 2.8% and 1.8% respectively), normalized to condition on at least one error: ¶ n ( ˆ u+ ˆ e = 1| FP a ( p)) = Binom(n, r c→u + r c→e )(1) 1− Binom(n, r c→u + r c→e )(0) (5) 14 Preprint. Under review. Forn∈1, 3, 5, 9, 13, we hand-label 20 sentences for each of two configurations:( ˆ e, ˆ u) = ( 1, 0)and(0, 1), weighting by their frequencies in the dataset to obtain¶ n (FP a ( p)| ˆ e + ˆ u = 1). Values for intermediatenare linearly interpolated. By design, all parameter choices yield upper bounds on false positive rates, ensuring our corrections are conservative. Adjustments to Counts.Letf (n)be the fraction of papers in a given subset withncitation sentences. The overall false positive rate for that subset is¶(FP a ( p)) = ∑ i f (i)· ¶ i (FP a ( p)) . Given total weight W = W c + W adopt , we adjust: W adopt ← W adopt − W· ¶(FP a ( p))(6) We do not directly correct for false negatives as they are negligible by construction. Because the base rate of adoption is low (roughly 0.41% of citing papers customize models), a 5% false positive rate among the much larger pool of context citations inflates adoption counts by∼54%. By contrast, a 5% false negative rate reduces the true count by only ∼5%. Additionally, a paper-level false negative requires all adoption sentences to be missed, whereas a single misclassified sentence suffices for a false positive. A.2.2 Sample Weighting Our text-extraction pipelines cover papers that are available online, preprint and open- access papers. To generalize estimates to the full population of citing publications, we construct inverse-probability weights using the complete S2AG citation graph. Fix a yeartand modelM. LetY M t be the total number of papers citingMin yeart(from S2AG) andX M t the number observed in our sample. A fraction of papers lack publication years; letX M 0 denote these, and letp t = X M t / ∑ τ̸=0 X M τ be the year distribution among dated papers. We define weights so that the observed sample, plus a proportional share of undated papers, recovers the population total: W M t = Y M t X M t + p t X M 0 t̸= 0 ∑ τ̸=0 p τ W M τ t = 0 (7) We apply analogous smoothing toY M t to handle missing years in S2AG. When aggregating to the paper level (across multiple cited models), we use the maximum sentence-level weight, which provides a lower bound on the true represented count by the union bound. A.3 List of Models • ALBERT • AraGPT2 Mega • BART • BLOOM 176B • ByT5-XXL • CPM-2 • CamemBERT • CodeGen Mono 16.1B • CodeT5 Base • CodeT5 Large • CodeT5+ • CogVLM • DeBERTa • FLAN-T5 • FLAN-UL2 • GLM-10B • GLM-130B • GPT 2 • GPT 3 • GPT 4 • GPT-NeoX-20B • Guanaco • InCoder • InstructBLIP 15 Preprint. Under review. • InstructGPT • Kosmos-2 • LLaMA • LLaMA 2 • LLaVA • LUKE • LongT5 • MT-DNN • MT0 • MiniGPT4 (Vicuna finetune) • NLLB • OPT • OPT-IML • PolyCoder • Pythia-12B • RoBERTa Large • SciBERT • StarCoder • T0-XXL • T5 • Tk-Instruct • Transformer-XL Large • WizardCoder • WizardLM • XLM • XLM-R XXL • XLM-RoBERTa • XLNet • mBART-50 • mT5-XXL B Supplemental Figures and Results Table 2: Formal inverted-U tests (full sample, 2019–2023). TestConditionEstimatep-value Quadratic regression Concavity ( ˆ β τ 2 < 0) ˆ β τ 2 =−0.041< 0.001 Peak within range ˆ τ † = 3.58∈ (0.5, 6.5)— Peak 95% CI (delta method)(3.30, 3.87)— Lind & Mehlum (2010) test Slope at τ min = 0.5 > 0+0.250t = 7.97 < 0.001 Slope at τ max = 6.5 < 0−0.236t = 6.19 < 0.001 Simonsohn (2018) two-lines test Rising phase (τ ≤ 3.58, n = 213) ˆ β = +0.123< 0.001 Falling phase (τ > 3.58, n = 57) ˆ β =−0.1010.036 All six conditions satisfied:✓ 16 Preprint. Under review. 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). SampleModelsObs ˆ β τ ˆ β τ 2 Peak (yrs)Peak (mos)R 2 Window Per-cohort 201913910.45 ∗ −0.06 ∗ 4.048.40.497 yr 20206360.56 ∗ −0.08 ∗ 3.440.30.386 yr 20216300.85 ∗ −0.15 ∗ 2.935.10.705 yr 202214561.05 ∗ −0.26 ∗ 2.024.30.534 yr 202319571.38 ∗ −0.45 ∗ 1.518.40.463 yr Pooled 2019–2022392130.36 ∗ −0.05 ∗ 3.845.60.28— 2019–2023582700.29 ∗ −0.04 ∗ 3.643.00.17— ∗ p < 0.001. HC1 robust standard errors. Peak =− ˆ β τ /(2 ˆ β τ 2 ). Figure 4: Adoption shapes for each model were manually classified into inverted-U, rising, plateau and unclear (noisy). Classification determines eligibility for time-to-peak and lifespan estimation. 0 500 1000 Count albert - inverted U-curve 0 1000 2000 bart-large - inverted U-curve 0 5000 Count gpt-4 - rising adoption Citations Adoptions 0 2000 4000 llama_2-7b - inverted U-curve 2019202020212022202320242025 Year 0 200 400 Count nllb - unclear 2019202020212022202320242025 Year 0 2000 4000 roberta_large - plateauing adoption Table 4: Sensitivity analysis for lifespan, varying threshold θ. θn ˆ β% per year 0.3049 −0.260 ∗ −22.9 0.4049 −0.260 ∗ −22.7 0.5049 −0.260 ∗ −22.6 0.6049 −0.250 ∗ −22.5 0.7049 −0.250 ∗ −22.5 ∗ p < 0.001. 17 Preprint. Under review. 20192020202120222023 Model release year 0 2 4 6 Time to peak * m (yrs) Later LLMs peak earlier (n=38) Change: -24%/year *** 20192020202120222023 Model release year Lifespan l m (yrs) LLM lifespan is shrinking (n=34) Change: -23%/year *** (a) Every model has at least four observed years (2022-2025 or longer). 20192020202120222023 Model release year 0 2 4 6 Time to peak * m (yrs) Later LLMs peak earlier (n=25) Change: -18%/year *** 20192020202120222023 Model release year Lifespan l m (yrs) LLM lifespan is shrinking (n=23) Change: -22%/year *** (b) Every model has at least five observed years (2021-2025 or longer). Figure 5: Sensitivity analysis. 2 4 6 Peak adoption age (yrs) Architecture decoder -31%/year *** encoder-decoder -30%/year *** Training type base -28%/year *** Open weights Open -27%/year *** 20192020202120222023 Model release year 2 4 6 Peak adoption age (yrs) Model size <10B -26%/year *** 10B -33%/year *** 20192020202120222023 Model release year Institution type academic -25%/year *** both -26%/year *** industry -29%/year *** 20192020202120222023 Model release year API access No API -27%/year *** API only -30%/year *** Peak Age Compression Varies by Model Characteristic Figure 6: Compression rate (time to peak) by model characteristic. 18 Preprint. Under review. Table 5: 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. % change/yr GroupnNo controls + log 10 params Panel A: Time to Peak τ ∗ All models55 −27% ∗ −30% ∗ Open53 −27% ∗ −30% ∗ No API52 −27% ∗ −30% ∗ Decoder28 −31% ∗ −32% ∗ Encoder11 −18%−20% Encoder-decoder16 −30% ∗ −30% ∗ Base37 −28% ∗ −32% ∗ Fine-tuned18 −22%−23% <10B params25 −26% ∗ −29% ∗ ≥10B params30 −33% ∗ −33% ∗ Academic9 −25% ∗ −26% ∗ Both17 −26% ∗ −25% ∗ Industry29 −29% ∗ −31% ∗ Panel B: Scientific Lifespan ℓ All models49 −23% ∗ −24% ∗ Open47 −23% ∗ −24% ∗ No API46 −23% ∗ −24% ∗ Decoder25 −24% ∗ −24% ∗ Encoder10 −16%−15% Encoder-decoder14 −24% ∗ −25% ∗ Base34 −23% ∗ −24% ∗ Fine-tuned15 −19% ∗ −20% ∗ <10B params22 −23% ∗ −25% ∗ ≥10B params27 −23% ∗ −23% ∗ Academic6 −19% ∗ −19% ∗ Both14 −22% ∗ −27% ∗ Industry29 −23% ∗ −24% ∗ Notes: Each cell reports the implied annual proportional change from ˆ β inln(DV) = α + β · release year + ε, estimated with HC3 robust standard errors.The right column adds log 10 (parameters)as a covariate.nreports the number of models in each subgroup. ∗ p <0.001; ∗ p < 0.01; ∗ p < 0.05. 19 Preprint. Under review. 2 4 6 Lifespan (yrs) Architecture decoder -23%/year *** encoder-decoder -23%/year *** Training type base -23%/year *** fine-tuned -19%/year *** Open weights Open -23%/year *** 20192020202120222023 Model release year 2 4 6 Lifespan (yrs) Model size <10B -23%/year *** 10B -23%/year *** 20192020202120222023 Model release year Institution type academic -19%/year *** both -22%/year *** industry -23%/year *** 20192020202120222023 Model release year API access No API -22%/year *** API only -25%/year *** Lifespan Compression Varies by Model Characteristic Figure 7: Compression rate (lifespan) by model characteristic. Table 6: Predictor analysis. Outcome: timetopeak. M1: SizeM2: + OrgM3: Full log(Param.)0.26**0.22*0.18 (0.10)(0.09)(0.12) API access0.13 (0.27) Fine-tuned-0.13-0.19 (0.19)(0.22) Encoder-0.46 (0.59) Encoder-decoder0.17 (0.33) Industry-0.08-0.09 (0.12)(0.19) Both (acad.+ind.)-0.25-0.19 (0.21)(0.23) Year fixed effects 2020-1.28**-1.32***-1.52*** (0.39)(0.39)(0.33) 2021-1.47***-1.47***-1.64*** (0.36)(0.37)(0.34) 2022-2.62***-2.59***-2.77*** (0.38)(0.41)(0.41) 2023-3.09***-2.99***-3.16*** (0.36)(0.34)(0.47) N555555 R 2 0.7970.8040.829 Year FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. 20 Preprint. Under review. Table 7: Predictor analysis. Outcome: lifespan. M1: SizeM2: + OrgM3: 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) Encoder-0.18 (0.35) Encoder-decoder-0.10 (0.20) Industry0.240.29 (0.15)(0.19) Both (acad.+ind.)0.230.28 (0.19)(0.22) Year fixed effects 2020-1.65***-1.60***-1.60*** (0.28)(0.28)(0.31) 2021-2.31***-2.24***-2.21*** (0.35)(0.37)(0.40) 2022-3.40***-3.23***-3.26*** (0.24)(0.29)(0.33) 2023-3.78***-3.70***-3.76*** (0.22)(0.24)(0.32) N494949 R 2 0.9200.9240.924 Year FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. 21 Preprint. Under review. Table 8: Predictor analysis. Outcome: logAdoptions. M1: SizeM2: + OrgM3: Full log(Param.)0.31**0.28*0.21 (0.10)(0.11)(0.13) API access0.70† (0.37) Fine-tuned-0.29-0.36† (0.18)(0.20) Encoder0.16 (0.36) Encoder-decoder0.10 (0.22) Industry0.43*0.34 (0.21)(0.22) Both (acad.+ind.)0.390.32 (0.25)(0.26) Year fixed effects 2020-0.95***-0.82***-0.88*** (0.28)(0.24)(0.24) 2021-1.28***-1.14***-1.16** (0.33)(0.32)(0.37) 2022-1.70***-1.37***-1.19*** (0.25)(0.31)(0.33) 2023-1.21***-1.00***-0.79* (0.26)(0.30)(0.34) N626262 R 2 0.3720.4380.478 Year FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. 22 Preprint. Under review. Table 9: Predictor analysis. Outcome: timetopeak. M1: SizeM2: + OrgM3: Full log(Param.)-0.40***-0.35**-0.38† (0.11)(0.12)(0.22) API access0.39 (0.47) Fine-tuned-0.54**-0.56* (0.19)(0.23) Encoder0.01 (0.67) Encoder-decoder0.10 (0.46) Industry0.50†0.44 (0.29)(0.40) Both (acad.+ind.)0.270.23 (0.32)(0.34) Year fixed effects Post 2022-1.30***-1.20***-1.12*** (0.18)(0.21)(0.30) N555555 R 2 0.4810.5530.559 Before/After 2022 FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. Table 10: Predictor analysis. Outcome: lifespan. M1: SizeM2: + OrgM3: Full log(Param.)-0.66***-0.55***-0.57* (0.13)(0.15)(0.24) API access0.22 (0.51) Fine-tuned-0.67*-0.60† (0.27)(0.31) Encoder0.04 (0.64) Encoder-decoder-0.19 (0.45) Industry0.97**0.98* (0.35)(0.42) Both (acad.+ind.)0.95*0.91* (0.41)(0.46) Year fixed effects Post 2022-1.50***-1.55***-1.54*** (0.23)(0.23)(0.32) N494949 R 2 0.5710.6620.667 Before/After 2022 FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. 23 Preprint. Under review. Table 11: Predictor analysis. Outcome: logAdoptions. M1: SizeM2: + OrgM3: Full log(Param.)-0.08-0.01-0.03 (0.10)(0.10)(0.15) API access0.78† (0.43) Fine-tuned-0.42*-0.40* (0.19)(0.20) Encoder0.30 (0.40) Encoder-decoder-0.04 (0.25) Industry0.74**0.63** (0.22)(0.23) Both (acad.+ind.)0.69*0.55* (0.27)(0.26) Year fixed effects Post 20220.040.020.14 (0.20)(0.22)(0.21) N626262 R 2 0.0120.2200.300 Before/After 2022 FE✓ † p <0.10,∗ p <0.05,∗ p <0.01,∗ p <0.001. HC1 robust standard errors. Table 12: 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). ModelYearArchitectureInstitutionRetention RoBERTa Large2019EncoderBoth0.79 GPT-2 (1.5B)2019DecoderIndustry0.78 DeBERTa2021EncoderIndustry0.72 CodeT5 Base2021Encoder-decoderBoth0.67 XLM-RoBERTa2019EncoderIndustry0.66 T5-3B2019Encoder-decoderIndustry0.59 T5-11B2019Encoder-decoderIndustry0.56 mT5-XXL2020Encoder-decoderIndustry0.54 SciBERT2019EncoderAcademic0.52 Transformer-XL2019DecoderBoth0.47 24