We sampled soils from a high plateau located in the upper belt of the Cordoba mountain grasslands in central Argentina (2100 m a.s.l. –above sea level, 31°34^′ S, 64°50^′ W). In this region, the mean temperatures of the coldest and warmest months are 5.1 and 11.5 °C, respectively, with no frost-free period. The mean annual precipitation is 900 mm, with most rainfall concentrated in the warmest months, between October and April. Soils are mostly Mollisols (Lithic Hapludolls), derived from weathering of granitic substrates and fine-textured eolian deposits (Cabido et al., 1987). The soil clays are dominated by biotite and illite, with a smaller proportion of kaolinite (Pasquini et al., 2002). The soil pH is 5.0 on average (Vaieretti et al., 2013).

We sampled soils from short grasslands, a vegetation community dominated by short annual and perennial grasses and forbs (e.g., Muhlenbergia peruviana (P. Beauv.) Steud. and Lachemilla pinnata (Ruiz & Pav.; Vaieretti et al., 2018, 2013). For this, we collected four compound soil samples (eight subsamples) from the 0–5 cm depth (4 replicates). Once collected, we sieved the soils through a 2 mm mesh, and we determined their soil water content using the gravimetric method.

For the incubation experiments, we placed 50 g of soil in 125 mL flasks. We applied two treatments: (1) control soils with no-litter addition, and (2) soils with litter-additions of 1 g of M. peruviana. The added litter was cut in small fragments (<2 mm) and mixed in the soil matrix. We selected M. peruviana because it is a dominant species in short grasslands, and also because of its high decomposability, in comparison to other grass species from the region (Poca et al., 2014; Vaieretti et al., 2013, 2018). We employed two time-sets of samples, as we destructively harvested them after 3 and 6 months of incubation (16 samples in total = 2 treatments × 2 incubation times × 4 replicates). This design also allowed us to have 8 replicates for respiration measurements until month 3, and 4 replicates until month 6. Throughout the incubation we maintained the soils at 25 °C and at field capacity (47 % water content; Cassel and Nielsen, 1986). To measure the soil respiration, we built closed microcosms where we placed the soils along a flask with water to avoid desiccation, placed CO₂traps using a flask with NaOH 1 M, and quantified the trapped CO₂ by titration.

We measured the accumulated respiration at 7, 15, 28, 42, 63, 91, 136 and 182 d after the beginning of the incubation. We measured the MAOC and POC contents of the soils before and after the incubations. For this we used the Duval et al. (2018) and Pestoni et al. (2020) techniques for MAOC and POC fractionation. Briefly, 10 g of air-dried soil < 2 mm was dispersed in 100 mL of distilled water and 10 glass beads (5 mm diameter) were added to increase aggregate destruction. The samples were subjected to mechanical dispersion through a rotary shaker (200 rpm) for 24 h. The soil suspension was poured through a 53 µm pore sieve using a water flow to separate the POC and MAOC fractions. These materials were washed into a dry dish, oven dried at 80 °C, and weighed. Then, we determined the MAOC, POC, and total C contents using the Walkley and Black technique (Nelson and Sommers, 1996). The fine (<53 µm) and coarse (>53 µm) fractions accounted for 74.3 % ± 2.2 % and 25.7 % ± 2.2 % of the recovered soil mass, respectively, and for 72.9 % ± 3.0 % and 27.1 % ± 3.0 % of the recovered C, respectively. Compared with bulk soil, fractionation resulted in high mass and C recoveries (97.52 % ± 0.37 % and 95.54 % ± 0.39 %, respectively).

Both differences in model structures and in parameter values might lead to different estimates of C persistence. To assess these effects, we examined how C persistence responds to (1) different model structures, (2) parameter changes, and (3) the combination of different model structures and parameter changes. For this, we ran simulations using the following 2- and a 3-pool models as a baseline, hereafter referred to as “2-Base” and “3-Base”, respectively (Fig. 1).

In the 2-Base model, new C enters the system directly into the POC pool and is subsequently transferred to the MAOC pool, as described in Eq. (8): 8dCdt=I0+-k10a2,1-k2POCMAOC, where I represents the C inputs in each time unit. In the 3-Base model, new C enters into the Litter C pool, is subsequently transferred to the POC pool, and finally to the MAOC pool, as described in Eq. (9): 9dCdt=I00+-k100a2,1-k200a3,2-k3LitterCPOCMAOC. To run the simulations, all the models received the same C input of 100 units per unit time. We applied the parameter values obtained from the optimization of the 3-pool models fitted to the litter-addition incubation treatments that did not assume a proportion of POC as Litter C (Table S4) across all model structures to ensure comparability. We selected these parameter values because they showed contrasting POC and MAOC decomposition rates and fell within the range of decomposition rates reported globally (Zhou et al., 2024). For the different model structures, we plotted the predicted probability density functions of C age and transit time and calculated their mean values. We used the following alternative model structures, shown in Table 1.

Over more than 80 years, and especially in recent decades, a wide variety of soil C models have been developed to describe and quantify SOC stocks and their persistence across a broad range of ecosystems and changing scenarios (Abramoff et al., 2018; Dangal et al., 2022; Manzoni and Porporato, 2009; Robertson et al., 2019; Zhang et al., 2021). While these models have proven to be highly useful in advancing our understanding of SOC dynamics (e.g., Georgiou et al., 2024; Gomes et al., 2019; Riggers et al., 2019; Zhang et al., 2024), their application still requires caution due to the implications of different model assumptions. In this study, focusing on the simple yet still widely used 2- and 3-pool models, we evaluated their ability to reproduce C pool and respiration dynamics using both experimental and simulation approaches, exploring whether the empirical results could be described by models with two or three distinct timescales.

Our findings indicate that 2-pool models were unable to capture both the C pool size changes and the respiration fluxes simultaneously. In contrast, 3-pool models were able to fit both response variables with a single set of parameters, yielding the lowest AIC and MSE values. It is important to highlight two key findings derived from the application of the models. First, it was almost self-evident that litter-addition treatments required a 3-pool structure, as fresh litter exhibits much faster decomposition dynamics than POC and MAOC. However, this was not expected for control soils (no litter-addition), which were presumably composed only of POC and MAOC. The fact that even for control soils the 3-pool models outperformed the 2-pool models might be because POC represented, indeed, a heterogeneous pool that should be modelled as distinct compartments. As a result, 2-pool models collapsed POC dynamics operating at different timescales into a single one, failing to capture both the fast initial and the slower later respiration phases simultaneously with the gradual C pool changes throughout the incubation.

Second, the 3-pool models showed a markedly better capacity to capture these faster and slower C dynamics for both variables. This suggests that the underlying mechanism in the inclusion of a third pool is enabling the model to distribute the processes operating at different timescales into the different C compartments. Specifically, the Litter C pool captured the rapid-response respiration fluxes at the beginning of the incubation, showing decomposition rates on the order of months. This allowed the POC and MAOC pools to represent intermediate and slower C dynamics, with decomposition rates on the order of years and decades, respectively (Table 2). In line with this, both C ages and transit times were consistently lower for the 3-pool models. This likely arises because these models estimate that the Litter C pool, which represents the C inputs to POC and MAOC, is rapidly respired and cycles quickly through the system. In contrast, in 2-pool models the respiration fluxes are driven by the longer timescale dynamics in the POC and MAOC pools, resulting in higher C persistence.

In recent decades, there has been a growing tendency to study SOC dynamics by conceptualizing SOC as POC and MAOC, fractions that usually reflect faster and slower dynamics, respectively. By definition, POC consists primarily of light-weight compounds of plant origin at different stages of decomposition, whose composition can vary with plant community and soil depth (Cotrufo et al., 2013; Lavallee et al., 2019; Von Lützow et al., 2007; Wiesmeier et al., 2019). This reflects varying proportions of recently added material and older, more processed C. As such, POC may contain a mixture of organic matter compounds with a wide range of decomposition and transfer rates. Moreover, recent studies have already employed different techniques to explore SOC heterogeneity, pointing out different functional C pools beyond POC-MAOC (e.g., Curtin et al., 2019; Leuthold et al., 2024). In this context, our results provide additional empirical and conceptual evidence that POC represents a heterogeneous pool containing compounds that cycle at different rates, which we hope will help improve current frameworks for understanding SOC dynamics. For our particular soils, POC's heterogeneity was a key driver of the observed differences in the model's predictions and performance. In light of our results, our goal is not to advocate for either 2- or 3-pool models, particularly given that several SOC models already employ a larger number of compartments (e.g., exchangeable and stable MAOC, free and occluded POC, dissolved organic C, microbial C; Abramoff et al., 2018; Manzoni et al., 2009; Witzgall et al., 2021; Zhang et al., 2021). Rather, our results emphasize that SOC models should explicitly represent processes operating across multiple temporal scales in ways that are appropriate for different systems and contexts, which may require combining experimental manipulations with modelling to evaluate model's performance. In our case, incorporating a third compartment beyond the POC-MAOC framework allowed us to reproduce the observed empirical dynamics, whereas in other contexts different model structures may be required.

When we simulated alternative models under different hypothetical scenarios and compared their estimates of C age and mean transit time to explore the effects of changes in parameters and model structure on C persistence, we observed that introducing a direct pathway to the MAOC pool (either through direct C inputs or through transfers) increased both C age and transit time. This response likely reflects that bypassing the relatively unstable Litter C and POC pools reduces early C losses, allowing a larger fraction of C to be retained in the more persistent and slower-cycling MAOC pool (Cotrufo et al., 2015; Kleber et al., 2015; Zhou et al., 2024).

It is important to note that, when only parameter values were modified, the changes in the Litter C decomposition rate had little effect on predicted C persistence. In contrast, increasing transfer rates led to modest increases in C age but much stronger increases in transit time. This pattern suggests that transfer processes, rather than Litter C decomposition rate alone, play a dominant role in transit times by controlling how rapidly C moves from labile pools into more persistent pools. Notably, when the transfer rates were tripled (Model 3.C and 3.D), the predicted C persistence was similar to or even higher than that obtained by adding a direct transfer pathway to MAOC (Fig. 5). This indicates that sufficiently high transfer rates can reproduce the effects of explicitly adding a direct transfer pathway to MAOC, as both mechanisms result in rapid C transfer into the MAOC pool.

When parameter and structural changes were combined, predicted C persistence increased substantially, particularly for transit time. In these scenarios, transit time values were approximately 7 times higher than those obtained for the baseline model, highlighting strong, non-additive interactions between model structure and key parameters (i.e., transfer rates into the MAOC pool). Overall, these results indicate that transfers into the MAOC pool have a stronger influence on the predicted C persistence. Therefore, when constructing models based on the POC-MAOC framework, assumptions regarding both model structure and transfer-related parameters should be carefully evaluated, as they can strongly influence estimates of C cycling and persistence. Model structure and transfer processes are particularly relevant in light of recent advances in our understanding of the factors that modulate transfers to the MAOC pool across ecosystems. For example, it is well established that MAOC formation efficiency is modulated by the saturation deficit, as well as by the presence of specific cations, such as oxalate-extractable Al and Fe and exchangeable Ca (Barré et al., 2014; Beare et al., 2014; Castellano et al., 2015; Saidy et al., 2013; Six et al., 2002). Hence, soils rich in these cations or with low saturation deficits may exhibit higher effective transfer rates to MAOC, which could translate into higher transit times. Vegetation composition differences might also play an important role, as rhizodeposition has been shown to promote higher MAOC formation efficiency than root or aboveground inputs (Villarino et al., 2021; Yin et al., 2025). In addition, recent evidence suggests that existing MAOC can promote the formation of new MAOC (King and Sokol, 2025), indicating potential feedbacks in transfer rates to MAOC. Despite these advances, further studies exploring MAOC formation are needed, as factors such as microbial carbon-use efficiency or C input rates have shown contradictory effects (King and Sokol, 2025; Sokol and Bradford, 2019; Wei et al., 2022; Yang et al., 2025).

Our results also suggest that transfer rates play a key role in determining SOC persistence. While current research has largely focused on quantifying SOC stocks and decomposition rates under different ecosystem contexts and management practices (e.g., Deng et al., 2016; Georgiou et al., 2024, 2022; Zhou et al., 2024), less attention has been given to explicitly quantifying transfer rates between C compartments. Future research should aim to characterize how C is transferred and transformed towards MAOC, and how these processes are represented in models. Reducing uncertainties associated with transfer rates among SOC compartments might be crucial for improving soil biogeochemical models and their ability to predict SOC persistence.