At present, there has been little investigation into the environmental effects of scientific and technological talent agglomeration. The overwhelming majority of the literature has examined the relationship between socioeconomic dynamics, such as industrial agglomeration and urbanization, on carbon emissions, arguing that these factors significantly enhance regional energy efficiency and environmental protection while highlighting the unintended negative consequences of neglecting a comprehensive approach (Zhang Y. et al., 2024). Some research shows that regional agglomeration of population and industry increases energy demand and carbon emission (York, 2007; Li et al., 2022). Nonetheless, when examining the agglomeration effects, researchers have largely found substantial evidence supporting improved public infrastructure efficiency, thereby reducing energy emission intensity (Glaeser and Kerr, 2009; Wuyts and Dutta, 2014).

Research supporting the first view contends that agglomeration can effectively reduce carbon emissions. This is because industrial agglomeration facilitates the scaling of regional industries, which helps lower production costs and promotes resource sharing, thereby improving energy efficiency and supporting environmental protection (Zou et al., 2024). Using panel data at the provincial level spanning 1991 to 2016, Zhang et al. (2017) demonstrated that regional industrial agglomeration can directly enhance regional production efficiency in the short term. High human capital not only promotes knowledge sharing and collaboration, driving process innovation and product quality improvement in enterprises (Dost and Badir, 2018), but also contributes to promoting sustainable practices through low-carbon policies, green technological innovation, and industrial optimization (Yang et al., 2022). High human capital promotes knowledge sharing and collaboration, significantly driving process innovation and product quality improvement in enterprises Regional industrial and population agglomeration has been recognized as a crucial strategy advancing environmental sustainability and fostering a green economy (Wang et al., 2020b).

In contrast, other research suggests that regional industrial and population agglomeration can exacerbate environmental degradation. On one hand, regional industrial and population agglomeration significantly increases energy demand, complicates the acquisition of resources such as land and energy, and leads to higher emissions of industrial pollutants (Andersson and Lööf, 2011). On the other hand, while agglomeration can improve accessibility in urban areas, it may also generate large volumes of waste and further raise energy consumption (Yi et al., 2022). For example, a study using urban population density and per capita energy consumption in China found a positive correlation between the two (Wang et al., 2024). Moreover, in highly agglomerated regions, the productivity gains associated with concentration may be accompanied by higher carbon emissions (Oliveira et al., 2014).

Nonetheless, certain research has suggested that the relationship between agglomeration economies and carbon emissions is nonlinear (Wang F. et al., 2020). Some researcher noted that there is significant difference across countries at different stages of development. While some scholars found that agglomeration in low-income countries effectively reduces carbon emissions, others observed the opposite trend in high-income countries, with an increase in carbon emissions. As demonstrated by Wang and Wang (2019), agglomeration is linked to environmental pollution in a manner that follows an inverted U-shaped trajectory. Initially, agglomeration generates sizable innovation and knowledge spillovers; once density surpasses a critical level, congestion and crowding erode these benefits, so the spillover gains weaken and may vanish. Only when surpassing a specific threshold does agglomeration begin to positively contribute to reducing pollution emissions (Wang F. et al., 2020; Wu et al., 2021). The research identified a positive N-shaped relationship between agglomeration and carbon emissions, suggesting that after surpassing an inefficient level of agglomeration, further agglomeration can theoretically mitigate carbon emissions (Li et al., 2022). Over extended periods, this relationship consistently exhibits a positive N shape, indicating that, theoretically, beyond certain levels of inefficiency, agglomeration can again contribute to reducing carbon emissions.

Despite the importance of talent agglomeration, research on this topic remains superficial, and the heterogeneous characteristics of these impacts have not been comprehensively explored. Therefore, this study further expands in three aspects: First, it establishes a bilateral analytical framework for assessing the effect of talent agglomeration on carbon emissions and evaluates this relationship across 30 provinces in China. Second, it investigates the twofold impact of talent agglomeration on regional carbon emissions and empirically explores its spatiotemporal evolution and changing patterns. Third, the study explores the differential impacts of urbanization and economic development at different levels of development and investigates practical paths to reinforce the inhibitory effects of talent agglomeration on emissions.

The agglomeration of elements within a region can enhance the quality and flow direction of these elements, thereby achieving economies of scale and generating economic benefits (Ye et al., 2022). But are the environmental impacts of spatial agglomeration the same? Agglomeration can promote the sharing of public facilities and technologies within a region, reduce pollution, and enhance collaborative innovation, thereby improving ecological efficiency (Tian et al., 2024). The increase in regional energy consumption and exacerbation of congestion caused by agglomeration place significant pressure on the sustainable development of the regional ecological environment (Martin and Sunley, 2006). As an essential prerequisite for innovation, scientific and technological talent extends urbanization and industrial agglomeration in regional development. Thus, its dynamics have two opposing effects on carbon utilization.

To measure the bilateral relationship between talent agglomeration and carbon emissions, we follow the two-tier stochastic frontier model of Kumbhakar and Parmeter (2009). As shown in Equation 1: Diff i t = x i t * δ + ω i t − u i t + ε i t = x i t * δ + ξ i t (1)

As shown in Equation 1, D i f f i t represents the carbon emission boundary, and   x i t denotes a set of control variables influencing regional carbon emissions, specifically urbanization, industrialization, energy structure and environmental regulation; ω i t represents the inefficiency component related to carbon emissions, u i t denotes the stochastic error term capturing other unobservable factors affecting carbon emissions. δ denotes the vector of parameters that require estimation The composite residual term ξ i t is defined ξ i t = ω i t − u i t + ε i t , where ε i t is a normally distributed random error term ε i t ∼ N 0 , σ 2 , accounting for deviations in the carbon emission gap.

Given that   ξ i t may not equal zero, relying on Ordinary Least Squares (OLS) estimates would result in bias. Therefore, the parameters are estimated using Maximum Likelihood Estimation (MLE) to ensure more accurate results. This model helps isolate the impact of talent agglomeration on the carbon emission gap, considering regional disparities in key economic and environmental factors. If ω i t ⩾ 0 , it indicates a positive effect of the concentration of technological talent on the carbon emission gap. Conversely, if u i t ⩽ 0 , it signifies a negative effect on the regional innovation efficiency gap, indicating a hindrance. When u i t ⩽ 0 , ω i t = 0 or ω i t ⩾ 0 , u i t = 0 , the model constitutes a unilateral stochastic frontier model.

If ω i t = u i t = 0 , the model reduces to an OLS model, which would be biased if ξ i t is non-zero. Therefore, MLE is applied to mitigate this bias and ensure more reliable estimates (Lei et al., 2024). The residuals   ε i t are assumed to follow a normal distribution with a mean of zero and variance σ ε 2 , i.e., ε i t ∼ i i d N 0 , σ ε 2 .

It is assumed that both ω i t and u i t are follow exponential distributions, specifically, ω i t ∼ i i d E X P σ ω , σ ω 2 and u i t ∼ i i d E X P σ u , σ u 2 , and u i t ∼ i i d E X P σ u , σ u 2 , with the residual terms being mutually independent and uncorrelated with inter-provincial characteristic factors. Given these distributional assumptions, the probability density function is derived, as shown in Equation 2: f ξ i t = exp α i t σ u + σ ω Φ γ i t + exp β i t σ u + σ ω ∫ − η i t ∞   φ x d x = exp α i t σ u + σ ω Φ γ i t + exp β i t σ u + σ ω φ η i t (2)

Within this framework, Φ · and φ · denote the cumulative distribution function (CDF) and the probability density function (PDF) of the standard normal distribution, respectively, and the reparameterizations are defined in Equation 3: α i t = σ v 2 2 σ u 2 + ξ i t σ u ; β i t = σ v 2 2 σ ω 2 − ξ i t σ ω γ i t = − ξ i t σ v − σ v σ u ; η i t = ξ i t σ v − σ v σ ω (3)

The log-likelihood to be maximized is given by Equation 4: ln ⁡ L X ; π = − n   ln ⁡ ⁡ σ ω + σ u + ∑ i = 1 n   ln ⁡ e α i t Φ γ i t + e β i t Φ η i t (4)

Within this framework, let π = β , σ v , σ ω , σ u . Consequently, the values for all parameters can be explicitly obtained via Maximum Likelihood Estimation (MLE). Furthermore, it is necessary to estimate ω i t and u i t , thus necessitating the further derivation of the conditional density function, as shown in Equations 5, 6: f ω i t ∣ ξ i t = 1 σ u + 1 σ ω exp ⁡ − 1 σ u + 1 σ ω ω i t Φ ω i t σ v + η i t exp ⁡ β i t − α i t Φ η i t + exp ⁡ α i t − β i t Φ γ i t (5) f u i t ∣ ξ i t = 1 σ u + 1 σ ω exp ⁡ − 1 σ u + 1 σ ω u i t Φ u i t σ v + η i t Φ η i t + exp ⁡ α i t − β i t Φ γ i t (6)

Based on Equations 5, 6, the conditional expectations follow from Equations 7, 8: E ω i t ∣ ξ i t = 1 1 σ u + 1 σ ω + σ v Φ − η i t + η i t Φ η i t exp ⁡ β i t − α i t Φ η i t + exp ⁡ α i t − β i t Φ γ i t (7) E u i t ∣ ξ i t = 1 1 σ u + 1 σ ω + exp ⁡ α i t − β i t σ v Φ − γ i t + η i t Φ γ i t Φ η i t + exp ⁡ α i t − β i t Φ γ i t (8)

The conditional expectations are used to quantify the absolute deviations of regional carbon utilization gaps from the frontier, representing positive and negative effects. For ease of subsequent discussion and comparison, these conditional expectations are further converted into percentages that represent deviations from the frontier level. Equations 9, 10 transform these expectations into percentage deviations from the frontier: E 1 − e − ω i t ∣ ξ i t = 1 − 1 σ u + 1 σ ω Φ γ i t + exp β i t − α i t exp σ v 2 2 − σ v η i t Φ η i t − σ v 1 + 1 σ u + 1 σ ω exp β i t − α i t Φ η i t + exp α i t − β i t Φ γ i t (9) E 1 − e − u i t ∣ ξ i t = 1 − 1 σ u + 1 σ ω Φ η i t + exp α i t − β i t exp σ v 2 2 − σ v γ i t Φ γ i t − σ v 1 + 1 σ u + 1 σ ω Φ η i t + exp α i t − β i t Φ γ i t (10)

Finally, the net effect is defined in Equation 11: N E = E 1 − e − ω i t ∣ ξ i t − E 1 − e − u i t ∣ ξ i t = E e − u i t − e − ω i t ∣ ξ i t (11)

Herein, N E represents the difference between the facilitating and hindering effects. If N E >0, it indicates that the facilitating effects are stronger than the hindering effects, thus playing a dominant role. Conversely, if N E <0, it signifies that the hindering effects are predominant.

Referring to the conceptual and empirical framework, and considering data availability, Chinese provincial panel data spanning 2013 to 2023 were used to investigate the effect of scientific and technological talent concentration on interprovincial atmospheric pollution. Owing to data constraints, the regions of Tibet, Hong Kong, Macao, and Taiwan were excluded from this study. All variables involving price factors were adjusted for inflation using 2010 as the base year.

Subsequently, we examine the spatial variation in the emission-reducing effects of talent agglomeration across China’s three major regions: Eastern, Central, and Western. Overall, talent agglomeration generates both promotive and inhibitory effects across regions, with the inhibitory effect consistently prevailing and leading to net emission reductions. These results suggest that while talent agglomeration consistently contributes to carbon emission reductions, the magnitude of its impact varies significantly by region.

In the western regions, where infrastructure and technological development levels are relatively low, the technological advancements and government policy support, including talent-rewarding policies and improvements in infrastructure, financial development, and credit optimization, driven by talent agglomeration can substantially improve carbon emission efficiency (Zhuo and Deng, 2020; Zheng et al., 2022). These measures strengthen innovation capacity and energy efficiency, ensuring that the inhibitory effect outweighs the promotive effect and produces the greatest emission reduction in the western region. In the central region, although energy-intensive industries still account for a larger share of the industrial structure (Fan et al., 2011), recent policy-driven talent agglomeration has enhanced efficiency improvements and management upgrading, leading to significant emission reductions (Yang et al., 2021). In the eastern regions, with better infrastructure and technological conditions, and a focus on technology industries and services, talent agglomeration primarily promotes management optimization and technological innovation (Chen et al., 2020). The overall evidence suggests that although both promotive and inhibitory effects coexist in all three regions, the inhibitory effect consistently dominates, producing regionally differentiated but uniformly negative net effects on carbon emissions.

Table 5 displays the frequency distributions of the promotive, inhibitory, and net effects of talent agglomeration on carbon emissions. The inhibitory effect exhibits a pronounced right-tailed distribution, extending beyond the 90th percentile, indicating that in certain provinces carbon emissions indicating that in certain provinces, large reductions occur as talent agglomeration increases. In contrast, the promotive effect weakens beyond the 90th percentile and remains below the inhibitory effect, underscoring its secondary role. The comparative distributions show that while both effects coexist, the inhibitory effect is more frequent and more substantial, consistently dominating the promotive effect. Accordingly, the overall distribution of effects is concentrated on the left half-axis below zero, with higher frequencies than on the right half-axis, confirming that the net impact of talent agglomeration is negative at most quantile points. These findings confirm the dominance of the inhibitory effect and underscore the need for region-specific emission reduction strategies.

Different from the linear or nonlinear regression methods used in other literature, the approach in this paper can be used to decompose the bilateral effects of talent agglomeration on carbon emissions for different years, and then analyze the trend characteristics of this impact over time. Figure 4 shows the annual variations in the promotive, inhibitory, and net effects. The results indicate that while both effects coexist, the inhibitory effect generally dominates, keeping the net effect negative in all years except 2014, when a temporary surge in promotion (46.27%) produced a positive net outcome (25.39%). The inhibitory effect peaked in 2013 (71.08%) and again in 2021 (67.80%), while the promotive effect reached its lowest in 2020 (19.90%). This produced a sizable negative net effect in 2020 (−17.36%), which deepened further by 2023 (−27.06%). These patterns suggest that while agglomeration may temporarily raise emissions in the early phase (e.g., 2014), its maturation strengthens the inhibitory channel through efficiency gains and structural upgrading, leading to persistently negative net effects that intensify after 2020.

In summary, the net effect of talent agglomeration on carbon emissions was predominantly negative during 2013–2023. Drawing on China’s practice, implementation exhibits a phased transition and lagged cumulative effects: an early and temporary scale-up (e.g., 2014), a near-neutral middle stage around 2018–2019, and a post-2020 efficiency-dominated phase in which the inhibitory channel strengthens with a lag (promotion bottoming in 2020 at 19.90%, inhibition peaking in 2021 at 67.80%). This shows that the impact of talent agglomeration on carbon emissions exhibits stage-specific dynamics and lagged materialization.

Drawing from the preceding mechanism analysis, this study hypothesizes that the effects of talent agglomeration on carbon emissions are twofold. The promotive effects may increase energy consumption and emissions through urban expansion, energy-intensive industrial structures and resource misallocation, while the inhibitory effects may enhance energy efficiency and reduce carbon intensity through knowledge exchange, green innovation and industrial upgrading. To examine the heterogeneous effects of talent agglomeration on carbon emissions, this study categorizes regions by urbanization, economic development, and environmental regulation levels, and tests the impacts across different percentiles. The results are presented in Table 6.

First, this study tested the effects of different degrees of talent agglomeration on emissions. In the 10%–40% range, emissions show a significant increase, followed by a decrease beyond 40%. This inverted U shape identifies a 40% agglomeration degree as the inflection point, demonstrating a substantial impact on sustainable development. However, regional disparities in China make it challenging for all areas to attract talent equally due to varying conditions for talent attraction.

Next, the study examined the urbanization level, a crucial element of the macroeconomic environment that is highly correlated with regional population structure and production methods (Nathaniel et al., 2021). At low levels of urbanization, the promotive effect dominates, as increased infrastructure construction and energy demand temporarily raise emissions. As urbanization advances, however, the inhibitory effect strengthens and overtakes the promotive effect, resulting in a net reduction in emissions. This pattern indicates that while both effects coexist, the balance shifts toward inhibition as urbanization deepens. In the 70%–100% range, the inhibitory effect becomes dominant, reducing emissions relative to lower levels of economic development. This further confirms the inverted U-shaped relationship, with promotive effects stronger at early stages and inhibitory effects prevailing as economic development advances.

Finally, we examined the relationship between talent agglomeration and carbon emissions under different levels of environmental regulation. At low levels of regulation (bottom 40%), the promotive effect dominates, and emissions increase. By contrast, in regions with stricter regulation, the inhibitory effect overtakes the promotive effect, resulting in significant net reductions. This further supports the view that emissions are more effectively reduced through environmental regulation than through other measures.

Three verification methods were adopted to ensure the robustness of the results.

First, considering that potential reverse causality may exist, in which talent agglomeration may affect carbon emissions while carbon emissions may also affect talent agglomeration, this paper adopts the instrumental variable (IV) approach to address the endogeneity problem. Based on this consideration, the first lag of talent agglomeration variable is selected as the instrumental variable, and a two-stage least squares (2SLS) estimation is conducted under the framework of province and year two-way fixed effects. The regression results are reported in Models M5 and M6 of Table 7. In M6, the Cragg–Donald Wald F statistic is 40.5, and the Kleibergen–Paap rk Wald F statistic is 35.2 (p < 0.001), indicating that the selected instrumental variable satisfies the relevance condition. As shown in M6, the impact of talent agglomeration on carbon emissions remains significantly negative, which is consistent with the main conclusions of the baseline regression and the bilateral frontier decomposition, suggesting that the findings of this paper are not driven by endogeneity due to reverse causality.

Then replacing the proportion of total city RCE to GDP as a proxy for carbon emissions still confirmed the existence of bilateral effects. The findings show that the promoting effect is 0.5590, while the inhibiting effect stands at 0.9486. This result confirms the statistical significance of talent agglomeration on carbon emissions and clarifies that talent agglomeration has bilateral effects on carbon emissions, consistent with the findings mentioned above.

Moreover, based on the impact weights, the promoting and inhibiting effects of talent agglomeration account for 9.2% and 90.8%, respectively. This indicates that over the years, the inhibiting effect has been dominant, which is consistent with the study’s conclusions. The variations in the inhibiting, net, and promoting effects were re-estimated, and these results are illustrated in Table 8. The study found that as the degree of talent agglomeration increases, its promoting effect raises carbon emissions, while the inhibiting effect curbs regional carbon emissions. The net effect leads to actual carbon emissions falling below the frontier level, aligning with the conclusions.

This study proposed a decomposition approach to examine the bilateral effects of talent agglomeration on carbon emissions. Utilizing panel data from 30 Chinese provinces from 2013 to 2023 and applying a two-tier stochastic frontier model, the study empirically decomposed the promotive and inhibitory effects of talent agglomeration and assessed their net impact. It further investigated the temporal, regional, and contextual heterogeneity of these effects.

The empirical findings reveal the following: (1) Talent agglomeration exhibits asymmetric bilateral effects on regional carbon emissions, with an overall inhibitory impact. On average, it increases emissions by 16.2% through the promotive channel but reduces them by 43.8% through the inhibitory channel, leading to a net reduction of 27.6% relative to the frontier level. (2) The net effect of talent agglomeration on carbon emissions is predominantly negative across different periods. After a brief increase in 2014, the inhibitory effect strengthened, particularly after 2020. (3) Regional heterogeneity is evident: the western region achieves significantly larger net reductions than the central and eastern regions. (4) Heterogeneity analysis shows that regions with higher talent agglomeration (≥40%), advanced economic development (70%–100%), and stricter environmental regulation experience greater net reductions in emissions, whereas at low urbanization stages the promotive effect dominates but shifts toward inhibition as urbanization deepens.

For policymakers aiming to harness talent agglomeration as a mechanism to promote sustainable development, our findings provide essential guidance, highlighting that while both promotive and inhibitory effects coexist, the inhibitory effects dominate overall, leading to a net reduction in carbon emissions. Tailoring strategies to fit specific regional conditions is therefore crucial to fully leverage talent’s role in emission reduction.

First, policymakers should adopt a demand-oriented approach to talent agglomeration, ensuring that talent attraction is closely aligned with local industrial development and environmental needs. Our findings demonstrate that enhancing the degree of talent agglomeration contributes significantly to effective emission reductions, if strategies are tailored to fit specific regional conditions. Accordingly, the identification and prioritization of key demands in industrial upgrading, R&D, and technology transfer is proposed, along with policies to guide the precise agglomeration of talent to address these needs. By aligning talent policies with local economic and environmental priorities, regions can foster multi-stakeholder collaboration and build sustainable innovation ecosystems that amplify the environmental benefits of talent agglomeration.

Second, the optimization of the spatial layout of talent agglomeration and the strengthening of cross-regional cooperation are advocated to enhance overall emission reduction. Empirical evidence shows that the inhibitory effect has strengthened since 2020, with the western region demonstrating greater emission reductions than the central and eastern regions. Regional cooperation should proceed along two complementary paths. Promoting local and neighboring talent exchange is anticipated to mitigate the negative spillover effects of over-agglomeration; meanwhile, cooperation and technical support with economically similar regions—through interregional talent cooperation funds, joint R&D, and technology transfer platforms—can unlock positive spillover potential and ensure that inhibitory effects dominate, resulting in net emission reductions. Ultimately, building cross-regional complementary innovation networks is essential to generate synergistic effects among regions, thereby enhancing national carbon reduction benefits and promoting high-quality economic development.

Finally, it is necessary to tailor talent agglomeration strategies to distinct regional conditions to strengthen the inhibitory effects on carbon emissions. Establish flexible, region-specific policies that align with the economic development stage, industrial structure, and existing talent levels in each area. Developed regions should consolidate the energy-efficiency gains of talent agglomeration, whereas less-developed regions should focus on building the foundations to enhance its inhibitory effects. In practice, this requires creating supportive platforms and career opportunities for talent retention—such as establishing region-specific priority projects and improving compensation schemes—so that talent agglomeration can more effectively contribute to carbon reduction and environmental sustainability across regions. These tailored strategies will ensure that talent agglomeration contributes effectively to carbon reduction and environmental sustainability in each region.

These findings open the door to future longitudinal research linking the bilateral effects of talent agglomeration on carbon emissions by showing that such agglomeration might be generally viewed as related to significant environmental outcomes, thus supplementing and extending previous studies. Like any other study, this study is not free from limitations. Owing to data constraints, this study analyzes the bilateral effects of talent agglomeration on carbon emissions from a macro perspective and does not account for the impact of micro-level enterprises and individual talent contributions on regional energy conservation and emission reduction. Ideally, these findings should be replicated in a study that explores a wider variety of environmental pollutants, including carbon dioxide emissions, wastewater contamination, and solid waste pollution. This approach would yield a more comprehensive understanding of how scientific and technological talent agglomeration impacts the environment.