For this systematic review and metanalysis, including GRADE qualification, we investigate the effects of MF on dynamic RE volume. We followed the methods established prior to the review and adhered to the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) for systematic reviews (Moher et al. 2009). Additionally, we prospectively registered the review in PROSPERO (ID: CRD420251003971).

The inclusion criteria for this systematic review are based on PICOS (Population, Intervention, Comparison, Outcome, and Study design) standards: (a) to be performed in healthy humans; (b) a high cognitive demands task intervention that requires sustained cognitive engagement, such as response inhibition, selective attention, or continuous performance monitoring; (c) a comparison group or experimental condition with low cognitive demands or passive control condition (e.g., watching documentaries and seated rest); (d) the outcome variable must be related to dynamic RE volume; and (e) a randomized controlled trial.

The search strategy is described in Figure 1. We searched the PubMed, Web of Science (WoS), SPORTDiscus, Scopus, and PsycINFO databases. The earliest available date up to November 2024 was considered in the process, with no language limitation. An updated search was conducted in February 2026 to ensure the inclusion of the most recent evidence. Search terms included combinations of MF and RE description [that is, (mental* fatigue* OR mental* effort* OR mental* exert* OR mental* exhaust* OR mental* strain* OR cognitive* fatigue* OR cognitive* effort* OR cognitive* exert* OR cognitive* exhaust* OR cognitive* strain* OR self‐control* depletion OR ego depletion) AND (resist* training OR resist* exercise OR strength* training OR strength* exercise OR muscular endurance* OR weight training OR weight exercise OR weightlifting OR power training OR power exercise OR circuit training OR circuit exercise OR plyometric training OR plyometric exercise)]. Search strategies were modified to meet the specific requirements of each database. The search was not limited to specific years. To minimize publication bias, we employed a gray literature search strategy encompassing: (a) backward citation searching of reference lists from all included studies and (b) systematic searching of the Open Science Framework (OSF), a free open‐source online platform hosting preprints and unpublished research. After identification, we first screened articles based on their titles and abstracts. Then, a full‐text evaluation was performed according to the PICOS criteria. Two researchers (LJFSJ and NA) performed the previous steps, and conflicts were solved by a third party (DLJ). For this review, RE was operationalized as isotonic (concentric‐eccentric) muscle contractions performed against an external resistance, including free weights, weight machines, resistance bands, and body weight. Isometric contractions and isokinetic testing protocols were excluded.

Each study's methods were evaluated using the TESTEX (Smart et al. 2015). Several tools are available to measure and report the quality of randomized controlled trials. However, in practice, some criteria, such as participant or researcher blinding, are sometimes unfeasible. In this case, TESTEX is the most appropriate tool to assess randomized controlled trials in physical exercise. The tool comprises 12 items, with scores ranging from 0 to 15: five points for quality and 10 for reporting. Each article was evaluated by two researchers (LJFSJ and GV), and a third researcher resolved disagreements.

The overall quality of the evidence for the outcome was evaluated using the Grading of Recommendations, Development and Evaluation (GRADE) system (Higgins et al. 2019). The following five factors were considered for classifying the quality of the evidence, where for each factor not met, the quality of evidence was reduced by one level (from high to moderate and low or very low) (Higgins et al. 2019): risk of bias (> 25% of the trials included in the comparison were classified as high risk of bias), inconsistency (I² > 50%), indirectness (> 50% of the participants were not related to the trial's target audience and differences in interventions), imprecision (< 400 participants in the comparison for continuous outcomes and > 300 participants for categorical outcomes), and publication bias (evaluated using a funnel plot when > 10 trials in the same comparison or an extensive search on databases and gray literature). Single‐trial comparisons (< 400 participants for continuous outcomes and < 300 participants for dichotomous outcomes) were considered to be inconsistent and imprecise, providing “low‐quality evidence”, which could be downgraded to “very low‐quality evidence” if limitations were identified regarding risk of bias (Higgins et al. 2019). The quality of the evidence was categorized as follows. Evidence was high quality if the results were consistent across ≥ 75% of participants, with low risk of bias, without publication bias, and with consistent, direct, and precise data; further research is unlikely to change the estimate or confidence in these results. Evidence was of moderate quality when one of the five classification factors above was not met; further research is likely to change the estimated effect and impact confidence in the effect. The evidence was of low quality because two of the five classification factors were not met; future research is likely to change the estimated effect and will have a significant impact on confidence in the effect. The evidence was very low quality when three of the five classification factors were not met; any estimate of effect is very uncertain.

The effect size for each study was calculated as the difference in performance (i.e., dynamic RE volume) between the experimental (i.e., MF) and control conditions. The mean difference and 95% confidence interval (CI) were calculated to measure the intervention effect on continuous outcomes. Standard deviations for the analysis were extracted using the methods described in the Cochrane Handbook (Higgins et al. 2019). The meta‐analysis was conducted using RevMan software (version 5.4) with a random‐effects model (standardized mean difference—SMD and Hedge's g adjusted for small sample bias) to calculate the pooled treatment effect with a 95% confidence interval as we anticipated that the selected studies varied in design, participants, and interventions, suggesting that these studies may not share the same true effect size. Thus, for between‐participants trials, we calculated Hedges' g (bias‐corrected SMD) using postintervention means and pooled SD. For within‐participants (crossover) trials, we computed standardized mean change with raw‐score standardization (SMCR) for the MF versus control difference. We converted it to Hedges' g, incorporating the within‐person correlation between conditions in the sampling variance (Morris and DeShon 2002). Missing means and standard deviations were calculated based on t‐values and p‐values for mean differences (Higgins et al. 2019) or obtained from figures using the WebPlotDigitizer (https://apps.automeris.io/wpd/). According to Hedge's g guidelines, SMD values of 0.2, 0.5, and 0.8 indicate small, moderate, and large effect sizes, respectively. Statistical significance was set at p ≤ 0.05. Statistical heterogeneity was assessed using the I² statistic and visual inspection of the forest plot.

To further characterize uncertainty in the context of between‐study heterogeneity, a 95% prediction interval (PI) was calculated for the primary meta‐analysis following the formula (Borg et al. 2024; Riley et al. 2011):

where:

Publication bias was systematically evaluated using multiple complementary methods, including visual inspection of funnel plot asymmetry, trim‐and‐fill adjustment procedure (Duval and Tweedie 2000a, 2000b), Egger's regression test (Egger et al. 2002), and Begg and Mazumdar's rank correlation test (Begg and Mazumdar 1994). The funnel plot serves as the primary graphical tool for detecting publication bias by examining the relationship between study precision (standard error) and effect size magnitude (J. A. Sterne and Egger 2001). To complement the visual assessment of funnel plot asymmetry, we conducted Egger's regression test, a formal statistical procedure that quantifies asymmetry through weighted linear regression. As a complementary nonparametric approach to assessing publication bias, we employed Begg and Mazumdar's rank correlation test (Begg and Mazumdar 1994), which evaluates the correlation between standardized effect sizes and their sampling variances using Kendall's tau rank correlation coefficient. Finally, to quantitatively estimate the potential impact of publication bias on the combined effect size, we employed the trim‐and‐fill method (Duval and Tweedie 2000a, 2000b), an iterative nonparametric procedure designed to: (1) identify and “trim” asymmetrically distributed studies, (2) re‐estimate the combined effect size based on the trimmed dataset, and (3) “fill” the funnel plot with imputed studies to achieve symmetry. Additionally, a test of excess significance (Ioannidis and Trikalinos 2007) was conducted to assess whether the number of statistically significant results among the included studies exceeded what would be expected given the estimated statistical power of the individual studies.

In the analysis, we identified 6396 studies, of which 5193 remained after duplicate removal. Following the screening of 5193 abstracts, 36 studies were selected for full‐text review and only 10 met the required standards. An updated search conducted in February 2026 identified additional records, from which two studies were selected for full‐text assessment. After careful evaluation against our eligibility criteria, one of these newly identified studies met all inclusion requirements. In total, 11 studies were included in the final analysis and 27 studies were excluded during the full‐text review stage (see Figure 1). The rationale for excluding each study is detailed in Supporting Information S1.

A detailed description of the studies is presented in Table 1. Study details and outcomes were divided into (a) study; (b) number of participants (N); (c) participants; (d) design; (e) cognitive task; (f) control task; (g) task duration; (h) manipulation check; and (i) warm‐up and physical task. All the data presented in Table 2 were independently extracted from studies by two review authors (LJFSJ and DLJ). Examination of funding sources, author affiliations, and conflict of interest statements revealed that all included studies reported no conflicts of interest.

More than 205 participants were included across the evaluated studies. The number of participants per study ranged from seven (Cuchna et al. 2024) to fifty (Graham et al. 2017). The participants consisted of physically active individuals (Brown et al. 2021; Graham et al. 2017; Head et al. 2016) and recreationally trained adults (Alix‐Fages, González‐Cano, et al. 2023a; Cuchna et al. 2024; De Lima‐Junior et al. 2024; De Salles Painelli et al. 2024; Gantois et al. 2021; Solon‐Júnior et al. 2026; Queiros et al. 2021). We also included hockey and rugby players (Dorris et al. 2012) and competitive rowers (Dorris et al. 2012). Therefore, this suggests low variability in the characteristics of the included subjects.

Regarding cognitive tasks, studies utilized social media on smartphone (Alix‐Fages, González‐Cano, et al. 2023a; Gantois et al. 2021), a congruent Stroop test (Cuchna et al. 2024; De Salles Painelli et al. 2024) and, primarily, an incongruent Stroop test (Alix‐Fages, González‐Cano, et al. 2023a; Brown et al. 2021; De Lima‐Junior et al. 2024; Graham et al. 2017; Solon‐Júnior et al. 2026; Queiros et al. 2021). Other studies used a challenging arithmetic/balance task (Dorris et al. 2012) or a vigilance test (Head et al. 2016). The duration of cognitive tasks varied significantly, ranging from five (Graham et al. 2017) to 52 min (Head et al. 2016). Only two studies reported an irregular time for cognitive tasks (Dorris et al. 2012; Solon‐Júnior et al. 2026).

Several manipulation checks were used to evaluate MF. Subjective measures were widely used; therefore, postcognitive effort data were collected from both groups (experimental vs. control). Subjective measures vary between studies, but the “M‐VAS” scale (i.e., MF visual analog scale) was widely used (Alix‐Fages, González‐Cano, et al. 2023a; Brown et al. 2021; De Lima‐Junior et al. 2024; De Salles Painelli et al. 2024; Gantois et al. 2021; Solon‐Júnior et al. 2026; Queiros et al. 2021). Other subjective measures, such as NASA‐TLX (i.e., Global workload and mental demand) (Cuchna et al. 2024; Head et al. 2016), ratings of perceived mental exertion (i.e., CR‐10 scale) (Graham et al. 2017), and a 7‐point scale (Head et al. 2016) were also used.

The selected studies were assessed using the TESTEX methodological quality scale. One study scored 9/15 (Cuchna et al. 2024), one study scored 10/15 (Dorris et al. 2012), two studies scored 11/15 (Graham et al. 2017; Head et al. 2016), four studies obtained 12/15 (Brown et al. 2021; De Salles Painelli et al. 2024; Gantois et al. 2021; Queiros et al. 2021), two studies scored 13/15 (Alix‐Fages, González‐Cano, et al. 2023a; Solon‐Júnior et al. 2026), and one scored 14/15 (De Lima‐Junior et al. 2024). Table 1 presents the detailed TESTEX scale scores for each study. The mean TESTEX score for the included studies was 11.58 ± 1.44 out of 15, indicating good methodological quality. To examine whether methodological quality influenced the pooled effect estimate, we conducted a sensitivity analysis excluding studies that scored ≤ 10/15 (≤ 67% of the maximum score), which represented the lower end of the quality distribution. This threshold‐based approach removed three comparisons from two studies (Cuchna et al. 2024; Dorris et al. 2012—both comparisons). The sensitivity analysis, including only studies with TESTEX > 10/15 (n = 9 studies and 11 comparisons), yielded results (g = −0.43; 95% CI: −0.63, −0.23; p < 0.01) that were similar to the primary analysis (g = −0.39; 95% CI: −0.56, −0.21; and p < 0.01), indicating that the findings are robust to variations in methodological quality.

The systematic search identified 14 interventions across 11 studies that examined the effects of MF on RE volume (number of repetitions or volume‐load). Although three studies (Alix‐Fages, González‐Cano, et al. 2023a; Dorris et al. 2012; Graham et al. 2017) contributed more than one comparison to the meta‐analysis, this approach was methodologically justified as each comparison involved distinct experimental conditions (e.g., different cognitive tasks or different participant populations within the same study). Therefore, this approach is consistent with Cochrane guidance for handling multiple intervention arms (Higgins et al. 2019). As shown in Figure 2, there was an overall effect of MF on resistance exercise volume [that is, number of repetitions and total training volume (TTV; number of repetitions × number of sets × load)] (standardized mean difference [SMD] = −0.39; IC 95% = −0.55 a −0.23; and p < 0.01), with significant heterogeneity (I² = 48% and p < 0.05). The 95% prediction interval for the overall effect ranged from −0.861 to 0.081, indicating that while the average effect is negative and statistically significant. This interval reflects the influence of between‐study heterogeneity (τ ² = 0.04 and I² = 48%) on the expected treatment effect in future applications and should be considered alongside the confidence interval when interpreting the practical relevance of these findings (Borg et al. 2024; Riley et al. 2011).

Visual inspection of the funnel plot revealed a relatively symmetric distribution with only a slight gap observed in the lower‐right side (Figure 3). Egger's regression analysis yielded a nonsignificant intercept (intercept = −3.88; 95% CI: −9.71, 1.94; t = −1.44; and p = 0.18), indicating that this test did not detect statistically significant asymmetry in the funnel plot. Similarly, the Begg and Mazumdar rank correlation test showed a weak inverse, nonsignificant relationship (Kendall's tau‐a = −0.23; z = −1.15; and p = 0.25) between the rankings of effect sizes and variances, providing no statistical evidence of publication bias using this nonparametric approach.

The Duval and Tweedie's trim and fill procedure was applied as a sensitivity analysis to estimate the number of potentially missing studies. The algorithm identified zero studies to be filled (L0 = 0). Consequently, the adjusted effect size remained identical to the observed estimate (Hedges' g = −0.39 and 95% CI [−0.56, −0.21]), indicating that the primary findings are robust against the “file drawer” effect. Finally, a test of excess significance (Ioannidis and Trikalinos 2007) was conducted to assess whether the number of significant results in the sample exceeded what would be expected based on the statistical power of the included studies. The observed number of significant results (O = 5) was highly consistent with the expected value (E = 5.52) (O/E ratio = 0.91), with a nonsignificant p‐value (p = 0.73). This lack of excess significance provides further evidence that the results have not been distorted by selective reporting of significant findings. However, given the small number of studies included in this meta‐analysis, caution is advised when interpreting these tests of small study effects and publication bias.

Subgroup analysis also showed significant differences between MF and control conditions for multijoint exercises [SMD = −0.45; 95% CI = −0.64, −0.26; and p < 0.01; (heterogeneity: p < 0.05 and I² = 53%)], but not for single joint exercises [SMD = −0.20; 95% CI = −0.43, 0.03; and p = 0.09; (heterogeneity: p = 0.71 and I² = 0%)] as shown in Figure 4.

For subgroup analysis by exercise intensity, loads were categorized according to established resistance training intensity zones (Carvalho et al. 2022): (a) low intensity: < 60% of one‐repetition maximum (1RM), encompassing very low and low (< 30%–59% 1RM) loads; (b) moderate intensity: 60%–79% 1RM; (c) high intensity: ≥ 80% 1RM; and (d) bodyweight exercises, in which external load relative to 1RM could not be determined. As shown in Figure 5, subgroup analysis revealed significant differences between MF and control conditions for very low intensity‐load [SMD = −0.40; 95% CI = −0.70, −0.11; and p < 0.01; (heterogeneity: p < 0.05 and I² = 60%)], moderate intensity‐load [SMD = −0.56; 95% CI = −0.81, −0.30; and p < 0.01; (heterogeneity: p = 0.10 and I² = 46%)], and bodyweight [SMD = −0.25; 95% CI = −0.50, 0.00; and p = 0.05; (heterogeneity: p = 0.14 and I² = 50%)]. The subgroup of high‐intensity load (≥ 80% 1RM) included only one study (De Lima‐Junior et al. 2024), which was insufficient to estimate a meta‐analysis. Accordingly, no pooled effect size is reported for this subgroup, and it is not included in subgroup comparisons.

Also, to examine whether the magnitude of mental fatigue effects varied according to total training volume characteristics, we conducted an additional subgroup analysis categorizing studies by relative training volume. The total training volume for each study was calculated as repetitions × sets × load intensity (%1RM). This calculation yields a relative volume metric representing cumulative training load. We could not incorporate absolute load (kilograms) into this calculation because most studies reported only the percentage of 1‐repetition maximum (%1RM) and the number of repetitions and sets completed, without reporting the actual loads lifted.

Based on the distribution of total training volume values across studies, we categorized comparisons into two subgroups: moderate total training volume (n = 5 comparisons from 5 studies) and high total training volume (n = 6 comparisons from 6 studies). The moderate total training volume subgroup demonstrated a significant negative effect of MF on resistance exercise performance (g = −0.37; 95% CI: −0.64, −0.10; and p < 0.01), with moderate heterogeneity (I² = 51% and p = 0.07). The high total training volume subgroup also showed a significant negative effect (g = −0.54; 95% CI: −0.84, −0.24; and p < 0.01), with moderate heterogeneity (I² = 52% and p = 0.08).

Based on the GRADE assessment (Figure 6), we observed low‐quality evidence (also referred to as low certainty of evidence) supporting the negative effect of mental fatigue on resistance exercise volume. The overall quality of evidence for each GRADE domain is presented in Figure 7. The evidence was downgraded for indirectness due to heterogeneity in mental fatigue induction methods across studies (e.g., Stroop test, smartphone‐based tasks, and vigilance tests) and variations in task duration (5–52 min), which likely contribute to the observed variability in effect magnitudes. Additionally, the evidence was downgraded for imprecision as the total number of participants included in the comparison (n = 205) fell below the optimal information size threshold of 400 participants recommended for continuous outcomes in GRADE assessments. This limited sample size reduces confidence in the precision of the pooled effect estimate.

All data generated or analyzed during this study are included in this published article (and its supplementary information files).