Effects of exercise on core symptoms in children with autism spectrum disorder: a systematic review and three-level meta-analysis Authors: Qingyuan Li, Zufeiya Tuerdi, Bo Li, Huiwen Guan, Wenqing Zhang, Yiyi Wang, Yanmeng He Published: 2026 ABSTRACT AbstractBackgroundThe prevalence of autism spectrum disorder (ASD) has increased substantially in recent years, yet evidence regarding the effects of exercise on core ASD symptoms remains inconclusive. Therefore, this study aimed to systematically evaluate the intervention and maintenance effects of exercise on core symptoms in children with ASD using a three-level meta-analysis, and to explore potential moderators and dose-response relationships.MethodsRandomized controlled trials (RCTs) examining the effects of exercise interventions on core symptoms in children with ASD were identified through searches of six databases: China National Knowledge Infrastructure (CNKI), PubMed, Web of Science, the Cochrane Library, Embase, and Scopus. Risk of bias was assessed using the Cochrane Risk of Bias tool 2.0 (RoB 2.0). Effect sizes were synthesized using the metafor package in R, with additional analyses including moderator analyses, dose-response modeling, sensitivity analyses, and assessments of publication bias. The certainty of evidence was evaluated using GRADEpro.ResultsA total of 20 studies involving 874 children with ASD were included. Moderate-certainty evidence indicated that exercise was associated with a statistically significant improvement in core ASD symptoms (Hedges’ g (g) = -0.53, 95% confidence interval (CI) -0.76 to -0.30, P < 0.001; 95% prediction interval (PI) -1.35 to 0.29). Low-certainty evidence suggested that the maintenance effects of exercise were not statistically significant (g = -0.46, 95% CI -0.97 to 0.06, P = 0.074; 95% PI -0.97 to 0.06). Age (F(1, 83) = 9.23, P = 0.003) and type of control condition (F(2, 82) = 4.80, P = 0.011) were identified as significant moderators of intervention effects, whereas core symptom domain, exercise modality, session duration, intervention duration, and exercise frequency were not significant moderators (P > 0.050). Furthermore, neither linear nor nonlinear spline models revealed significant associations between total exercise volume or weekly exercise duration and intervention effects (P > 0.050).ConclusionsBased on the current evidence, exercise interventions may be associated with improvement in core symptoms in children with ASD, with effects moderated by age and the type of control condition. However, these findings should be interpreted cautiously in light of the prediction interval crossing zero, the methodological limitations of the included studies, and the limited evidence regarding maintenance effects. No significant dose-response relationship was observed for total or weekly exercise duration. Overall, exercise may be considered a promising supportive approach within the broader range of interventions for ASD, rather than a stand-alone or definitively established strategy. Its long-term effects and optimal exercise prescriptions still require confirmation through high-quality studies.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12888-026-08319-z.Keywords: Exercise, Autism spectrum disorder, Children, Core symptoms, Meta-analysis Introduction Autism spectrum disorder (ASD), also referred to as autism, is a neurodevelopmental disorder with onset in early childhood. In recent decades, the prevalence of ASD has shown a continuous upward trend, making it a major public health and educational concern worldwide. According to the most recent data released by the U.S. Centers for Disease Control and Prevention (CDC) based on its surveillance network, the prevalence of ASD among 8-year-old children has reached nearly 1 in 31, indicating a growing disease burden [1]. Systematic reviews and meta-analyses further suggest substantial variability in ASD prevalence across countries and regions, while consistently highlighting an overall increasing trend and high societal burden, underscoring the importance of early identification and intervention [2]. According to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), the core clinical features of ASD are characterized by: (1) persistent deficits in social communication and social interaction (social communication and interaction, SCI), and (2) restricted, repetitive patterns of behavior, interests, or activities (restricted and repetitive behaviors, RRB) [3]. In addition to these core features, many children with ASD may also experience difficulties in language development, sensory processing, adaptive functioning, and executive functioning [4, 5]. Collectively, these difficulties substantially affect children’s learning abilities, peer relationships, and family functioning, and may have profound long-term consequences for their developmental trajectories across the lifespan [5]. Previous studies have estimated that the lifetime societal costs associated with ASD can reach the order of millions of U.S. dollars and increase markedly with symptom severity and support needs; in individuals requiring high levels of support, lifetime costs have been estimated to range from approximately 1.4 to 3.6 million USD, placing sustained pressure on families and social resources [6]. Therefore, identifying safe, sustainable, and scalable intervention strategies to improve core ASD symptoms and related functional impairments remains a critical priority. Exercise may represent a promising intervention approach not only because of its accessibility, low cost, and adaptability, but also because it may influence several processes closely related to core ASD symptoms. Emerging evidence suggests that physical activity may enhance executive and attentional control, promote sensorimotor integration, support neuroplasticity, and provide structured opportunities for social engagement and behavioral practice. These changes may, in turn, facilitate social communication and interaction while helping to reduce behavioral rigidity and restricted or repetitive behaviors [7–10]. These benefits may be particularly relevant because many structured exercise programs involve turn-taking, imitation, rule-following, and shared engagement, which are closely related to social communication and behavioral regulation in children with ASD. In recent years, exercise-based interventions have received increasing attention due to their low cost, high accessibility, and adaptable adherence. A growing body of research has examined the effects of exercise on core symptoms in children with ASD. For example, Toscano et al. implemented a 24-week moderate-intensity exercise program (two sessions per week, 30 min per session) in children with ASD and reported improvements in social interaction and attention-related outcomes [11]. In contrast, Howells et al. conducted a football-based intervention delivered once per week for 60–90 min and found no significant improvements in social communication or language outcomes in children with ASD. Such inconsistent findings may be attributable to heterogeneity in exercise protocols as well as individual differences among children with ASD [12]. Several meta-analyses have attempted to synthesize evidence on the effects of exercise on core ASD symptoms. Wang et al. included 16 studies and conducted a conventional meta-analysis, reporting overall beneficial effects of exercise on core symptoms; however, subgroup analyses suggested that interventions lasting less than 12 weeks, with session durations of 60–90 min or exercise frequencies of one to two sessions per week, did not yield significant improvements [13]. Similarly, Cai et al. synthesized eight studies using a traditional meta-analytic approach and found that exercise improved core symptoms in individuals with ASD, although potential moderating factors were not further explored [14]. Despite these efforts, the existing evidence base on exercise interventions for core symptoms in children with ASD has several important methodological and empirical limitations. First, individual exercise trials frequently assess multiple core-symptom-related outcomes within the same study, such as social communication abilities, repetitive behaviors, and attention or emotional problems [15–17]. However, most previous meta-analyses have relied on conventional two-level models, typically selecting a single outcome per study or combining multiple outcomes using simplified approaches. Such practices ignore the statistical dependence among effect sizes within the same study and may lead to information loss or underestimated standard errors, thereby compromising the robustness of effect estimates [18, 19]. Methodological research has demonstrated that three-level meta-analysis can explicitly model both within-study and between-study variability when multiple correlated effect sizes are present, allowing more efficient use of available data and improved statistical power [18]. Nevertheless, this approach has not yet been systematically applied in the context of exercise interventions targeting core ASD symptoms. Second, evidence regarding the maintenance effects of exercise interventions remains limited. Existing systematic reviews and meta-analyses consistently indicate that most studies focus on immediate post-intervention outcomes, with insufficient follow-up assessments, leaving the long-term sustainability of exercise effects on core ASD symptoms largely unclear [20, 21]. Given that ASD is a lifelong neurodevelopmental condition, the durability of intervention effects is of particular relevance for both clinical and educational practice, yet this issue has not been adequately addressed in the current evidence base. Finally, substantial heterogeneity exists across studies with respect to exercise modality, intervention duration, session length, weekly frequency, and exercise intensity, and findings across studies remain inconsistent [22, 23]. However, previous meta-analyses have provided only limited examination of potential moderators and dose-response relationships. In particular, whether total exercise volume and weekly exercise duration are associated with improvements in core symptoms in a linear or nonlinear manner remains unclear due to a lack of systematic quantitative evidence. Therefore, the present study aims to apply a three-level meta-analytic model to comprehensively evaluate the immediate intervention effects and maintenance effects of exercise on core symptoms in children with ASD, while maximizing the inclusion of all relevant effect sizes related to SCI, RRB, and overall core symptom indices. Furthermore, moderator analyses and both linear and nonlinear spline dose-response models will be conducted to examine the associations between key exercise characteristics (e.g., exercise modality, intervention duration, frequency, and session length), exercise dose (weekly exercise duration and total exercise volume), and intervention effects. This study seeks to provide more targeted and actionable evidence to inform the optimization of exercise-based intervention programs for children and adolescents with ASD. Method Study design and registration This systematic review and meta-analysis was conducted and reported in accordance with the methodological guidance of the Cochrane Handbook for Systematic Reviews of Interventions. The study protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO) under the registration number CRD420251230426. Eligibility criteria Eligibility criteria were defined according to the Population, Intervention, Comparison, Outcome, and Study design (PICOS) framework. Population (P): Participants were required to have a formal clinical diagnosis of ASD based on established diagnostic classification systems, such as the DSM (DSM-IV or DSM-5) or the International Classification of Diseases (ICD). When reported, standardized assessment instruments such as the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) were considered supportive tools used within the diagnostic process rather than independent diagnostic classification systems. All included studies were required to report that ASD diagnoses had been established or confirmed by licensed clinicians, multidisciplinary diagnostic teams, or other formally recognized diagnostic services. Studies recruiting participants from educational settings were included only when the articles explicitly indicated that participants had previously received a formal ASD diagnosis or had undergone standardized diagnostic evaluation procedures. Only studies involving children aged 4 to 12 years were included. This age range was selected to focus on a developmentally comparable period in which ASD diagnosis is relatively established, core symptom measures are more consistently applied, and exercise-based interventions are more comparable across studies. In addition, most eligible randomized controlled trials (RCTs) identified in the present review were conducted within this age range, which helped reduce developmental heterogeneity across included samples. Intervention (I): The intervention of interest was exercise, defined as any structured or unstructured physical activity, including but not limited to yoga, basketball, kata, and horseback riding. Studies in which exercise was combined with other components were included only when exercise constituted a core element of the intervention and the non-exercise component was also provided to the comparison group or was unlikely to account for the main treatment effect independently. Comparison (C): Control conditions included usual care, muscle relaxation, wait-list control, or other non-exercise-based interventions. Outcome (O): According to the DSM-5, the two core symptom domains of ASD are SCI and RRB. Commonly used outcome measures included the Gilliam Autism Rating Scale, Second Edition (GARS-2), Social Responsiveness Scale (SRS), Sensory Profile, Autism Treatment Evaluation Checklist (ATEC), and Aberrant Behavior Checklist (ABC). Study Design (S): Only RCTs were eligible for inclusion. Exclusion criteria Studies were excluded if they were non-randomized trials; if relevant data could not be extracted; if participants’ ages did not meet the inclusion criteria; if publications were duplicates; or if they were reviews, conference abstracts, or other non-original research articles. Literature search strategy and study selection A comprehensive computerized literature search was conducted in six databases, including China National Knowledge Infrastructure (CNKI), Web of Science(WOS), PubMed, the Cochrane Library, Embase, and Scopus, from database inception to November 14, 2025. Medical Subject Headings (MeSH) terms and free-text terms were combined using the Boolean operators “OR” and “AND.” The primary search terms included “physical activity,” “exercise,” “training,” “sport,” “ASD,” “autism spectrum disorder,” and “stereotyped behavior.” In addition, reference lists of relevant systematic reviews and meta-analyses were manually screened to identify additional eligible studies. The detailed search strategies for each database are provided in Supplementary Table S1. All retrieved records were imported into EndNote X9, and duplicate publications were removed. Two reviewers independently screened titles and abstracts to exclude irrelevant studies. Full texts of potentially eligible articles were then assessed independently for final inclusion. Any disagreements between the two reviewers were resolved through discussion with a third reviewer until consensus was reached. Data extraction and coding Two reviewers independently extracted data from all included studies using a predefined data extraction form. Extracted information covered three main domains: study characteristics, baseline comparability, intervention characteristics, and outcome measures. Study characteristics included the first author’s name, year of publication, country, total sample size, and proportion of female participants. Intervention characteristics comprised exercise modality, session duration, intervention duration, exercise frequency, and the content of the control condition. Whether the intervention was exercise-only or multi-component was also recorded for sensitivity analysis. Outcome measures focused on data related to core ASD symptoms. For each study, change scores were extracted for both the intervention and control groups. Change scores were defined as the difference between post-intervention and baseline values (post-pre). In the present meta-analysis, all pooled effect sizes were derived from between-group differences in pre-post change scores rather than post-intervention values alone. This approach was adopted to reduce the potential influence of baseline between-group differences on the estimated intervention effects. In addition, exploratory analyses were conducted to examine the maintenance effects of exercise; for this purpose, change scores from post-intervention to follow-up assessments were also calculated when follow-up data were available. If a study included two or more intervention groups or control groups, all eligible groups were extracted separately and included in the analysis. When required data (e.g., means or standard deviations) were not reported, the corresponding authors were contacted to obtain additional information.For continuous outcomes, the mean change score was calculated using the following formula (post–pre): Mean_change=Mean_post-Mean_baseline. If the standard deviation of the change score (SDchange) was not reported, it was estimated according to the recommendations of the Cochrane Handbook using baseline and post-intervention standard deviations: where r represents the correlation coefficient between baseline and post-intervention measurements. Because this coefficient was not reported in most included studies, we followed the recommendations of the Cochrane Handbook and common practice in meta-analyses of continuous outcomes and assumed r = 0.50 as a conservative mid-range estimate when pre-post correlations were unavailable [24, 25]. Any disagreements during data extraction and coding were resolved through discussion with a third reviewer until consensus was reached. Moderator analyses were conducted in this study, and the coding scheme was defined a priori as follows. Core symptom domains were coded as SCI and RRB. Exercise modality was categorized into four groups: traditional fitness/skill exercise, martial arts-based exercise, aquatic/sensory-based exercise, and mind-body exercise. Control conditions were coded as passive control, usual care control, and active control. Participant age was categorized as preschool children and school-aged children. Session duration was coded as < 60 min or ≥ 60 min. Intervention duration was categorized into 4–9 weeks, 10–12 weeks, and > 12 weeks. Exercise frequency was coded as < 3 sessions per week or ≥ 3 sessions per week. These moderator categories were defined a priori based on the distribution of the included studies, commonly used categorizations in previous exercise-related meta-analyses, and the need to retain an adequate number of effect sizes within each subgroup for stable estimation. In addition to the pre-specified age subgroup analysis, age was further examined as a continuous moderator in a three-level meta-regression model to assess whether intervention effects varied linearly with mean age across studies. Detailed coding information is provided in the Supplementary Materials. In addition, dose-response relationships between exercise exposure and intervention effects were examined. Weekly exercise duration was calculated by multiplying session duration by exercise frequency, and total exercise volume was calculated by multiplying weekly exercise duration by the total number of intervention weeks. All time variables were converted into minutes for consistency. When session duration was reported as a range (e.g., 30–60 min), the midpoint of the interval was used for calculation. To reduce reliance on categorical cut-points, weekly exercise duration and total exercise volume were also analyzed as continuous variables in the dose-response models. Risk of bias and certainty of evidence assessment The internal validity of the included RCTs was systematically assessed using the revised Cochrane Risk of Bias tool for randomized trials (RoB 2.0), as recommended by the Cochrane Collaboration [26]. The RoB 2.0 tool evaluates the risk of bias across five domains: (1) bias arising from the randomization process; (2) bias due to deviations from intended interventions; (3) bias due to missing outcome data; (4) bias in measurement of the outcome; and (5) bias in selection of the reported result. Each domain was judged based on a series of signaling questions and classified as “low risk,” “some concerns,” or “high risk of bias.” An overall risk-of-bias judgment for each study was then derived by integrating the assessments across all five domains. The certainty of evidence for the primary outcomes was evaluated using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach. The GRADE framework considers five factors: risk of bias, inconsistency, indirectness, imprecision, and publication bias, and classifies the certainty of evidence into four levels: high, moderate, low, and very low. RCTs were initially rated as high-certainty evidence and were subsequently downgraded or upgraded based on these domains to determine the overall confidence in the pooled estimates [27]. All assessments were conducted independently by two reviewers. Any disagreements were resolved through discussion with a third reviewer until consensus was reached. Statistical analysis Three-level meta-analyses were conducted using the metafor package in R [28–31]. Effect sizes were expressed as Hedges’ g (g) and were calculated based on between-group differences in pre-post change scores rather than post-intervention values alone. Corresponding 95% confidence intervals (CIs) and 95% prediction intervals (PIs) were calculated. Based on previous literature, effect sizes were interpreted as small (g < 0.20), small to moderate (g = 0.20–0.49), moderate (g = 0.50–0.79), and large (g ≥ 0.80). Overall heterogeneity was assessed using the Q statistic. Variance components were estimated at three levels, including sampling variance (Level 1), within-study variance (Level 2), and between-study variance (Level 3). The statistical significance of each variance component was evaluated using one-sided log-likelihood ratio tests (LRT), as recommended for three-level meta-analytic models [30]. To examine the robustness of the pooled effect estimates, leave-one-out (LOO) sensitivity analyses were performed, and influence diagnostics were calculated using influence analyses. If omitting any single study resulted in a change in the direction or statistical significance of the pooled effect size, or if a study exhibited notably high residuals or influence values, that study was considered to have a potential influence on the overall results. Otherwise, the pooled findings were regarded as robust [28]. Additional sensitivity analyses were conducted to examine the robustness of the findings by excluding (1) effect sizes with significant baseline differences, (2) studies with insufficient reporting of baseline comparability, and (3) multi-component interventions, in order to assess the potential influence of baseline imbalance and combined intervention designs on the pooled estimates. To investigate dose-response relationships between exercise exposure and intervention effects on core ASD symptoms, both linear models and nonlinear spline models were fitted within the three-level meta-analytic framework [32]. In addition, potential publication bias was examined using Egger’s regression test, visual inspection of funnel plots, and the trim-and-fill method [33, 34]. Results Study selection A total of 1,778 records were identified through database searches. After removing 697 duplicate records, 1,081 articles remained for title and abstract screening. Of these, 1,061 records were excluded as irrelevant. The full texts of the remaining 65 articles were retrieved and assessed for eligibility. Among these, 14 studies were excluded because their outcome measures did not meet the inclusion criteria, 7 were reviews or conference abstracts, 2 had unavailable or non-extractable data, 1 was a duplicate publication, and 21 were non-RCTs. Ultimately, 20 RCTs were included in the final analysis [35–54].The study selection process is illustrated in Fig. 1. Flow diagram of study selection Characteristics of included studies and interventions Of the 20 included studies, eight were conducted in China, six in Iran, three in the United States, and one each in Tunisia, India, and Australia. In total, 874 children with ASD were included, with ages ranging from 4.06 to 11.93 years. Exercise interventions encompassed a variety of modalities, including kata techniques training, basketball, horseback riding, yoga, and Nei Yang Gong. Session duration ranged from 30 to 90 min, exercise frequency varied from one to five sessions per week, and intervention duration ranged from 4 to 24 weeks. Control conditions included usual care, wait-list control, and maintenance of daily activities. Five studies reported follow-up assessments. One study conducted follow-up assessments two months after the completion of the exercise intervention [35], while four studies conducted follow-up assessments one month post-intervention [36, 37, 42, 45]. Outcome measures for core ASD symptoms included the ABS, the SRS, and the GARS-2, among others. With regard to baseline comparability, 16 studies reported no significant baseline differences between groups, baseline reporting was insufficient for some outcomes in 2 studies, and 2 studies reported significant baseline differences. Detailed information on study characteristics and intervention features is presented in Table 1. Table 1Characteristics of included studies and intervention featuresFirst author (year)CountrySample size/Age/Percentage of femalesBaseline comparabilityIntervention featuresOutcomeArabi et al. 2019Iran60/8.51/25%(1)Exercise modality: Visuomotor / motor;Session duration: 50 min;Intervention duration: 10 weeks/Two months follow up;Exercise frequency: 3 days a week;Exercise intensity: Unclear;Control: Visual training/ Routine care.ADOSBahrami et al. 2012Iran30/9.13/13.33%(1)Exercise modality: Kata techniquestraining;Session duration: 30–90 min;Intervention duration: 14 weeks/One month follow up;Exercise frequency: 4 days a week;Exercise intensity: Unclear;Control: No-exercise control.GARS-2Bahrami et al. 2016Iran30/9.13/13.33%(1)Exercise modality: Kata techniquestraining;Session duration: 30–90 min;Intervention duration: 14 weeks/One month follow up;Exercise frequency: 4 days a week;Exercise intensity: Unclear;Control: No-exercise control.GARS-2Bass et al. 2009USA34/7.34/14.71%(2)Exercise modality: Therapeutic horseback riding;Session duration: 60 min;Intervention duration: 12 weeks;Exercise frequency: 1 day a week;Exercise intensity: Unclear;Control: Wait-list control.SRS, SPChan et al. 2013China40/11.85/20.00%(1)Exercise modality: Nei Yang Gong;Session duration: 60 min;Intervention duration: 4 weeks;Exercise frequency: 2 days a week;Exercise intensity: Unclear;Control: Progressive muscle relaxation.ATECGabriels et al. 2015USA116/10.20/12.93%(1)Exercise modality: Therapeutic horseback riding;Session duration: 45 min;Intervention duration: 10 weeks;Exercise frequency: 1 day a week;Exercise intensity: Unclear;Control: Barn activity control intervention.SRS, ABCHaghighi et al. 2023Iran16/8.57/43.75%(1)Exercise modality: Combined physical training;Session duration: 60–80 min;Intervention duration: 8 weeks;Exercise frequency: 3 days a week;Exercise intensity: Unclear;Control: Did not perform any regular physical training.GARS-2Ju et al. 2024China17/11.93/29.42%(1)Exercise modality: Yoga;Session duration: 45–50 min;Intervention duration: 8 weeks/One month follow up;Exercise frequency: 3 days a week;Exercise intensity: Unclear;Control: Daily program activities.ABCLindor et al. 2023Australia27/9.04/51.85%(1)Exercise modality: AllPlay Dance;Session duration: 60 min;Intervention duration: 8 weeks;Exercise frequency: 1 day a week;Exercise intensity: Unclear;Control: Wait-list control.SRSMarzouki et al. 2022Tunisia22/6.33/9.09%(1)Exercise modality: Technical aquatic training/Game-based aquatic training;Session duration: 50 min;Intervention duration: 8 weeks;Exercise frequency: 2 days a week;Exercise intensity: Unclear;Control: Daily program activities.GARS-2Movahedi et al. 2013Iran30/9.13/13.33%(1)Exercise modality: Kata techniquestraining;Session duration: 30–90 min;Intervention duration: 14 weeks/One month follow up;Exercise frequency: 4 days a week;Exercise intensity: Unclear;Control: No-exercise control.GARS-2Liu et al. 2021China23/10.10/21.74%(2)Exercise modality: Gross motor intervention;Session duration: 60 min;Intervention duration: 6 weeks;Exercise frequency: 4 days a week;Exercise intensity: Moderate;Control: Daily program activities.ASSS, SRSDong et al. 2020China30/4.82/16.67%(3)Exercise modality: Mini-basketball Training;Session duration: 40 min;Intervention duration: 12 weeks;Exercise frequency: 5 days a week;Exercise intensity: Moderate (128–148 beats/minutes);Control: Routine rehabilitation.RBSLiu et al. 2024China72/4.96/29.17%(1)Exercise modality: Group sports game;Session duration: 60 min;Intervention duration: 12 weeks;Exercise frequency: 5 days a week;Exercise intensity: Unclear;Control: Routine rehabilitation.SRSZhang et al. 2025China100/5.13/36.00%(1)Exercise modality: Sports games combined with music therapy;Session duration: 30 min;Intervention duration: 12 weeks;Exercise frequency: 5 days a week;Exercise intensity: Unclear;Control: Music therapy.ATECPhung et al. 2021USA34/9.31/17.65%(1)Exercise modality: Mixed martial arts training;Session duration: 45 min;Intervention duration: 13 weeks;Exercise frequency: 2 days a week;Exercise intensity: Unclear;Control: Wait-list control.SSISJafari et al. 2025Iran54/4.06/Unclear(1)Exercise modality: Basketball/Karate;Session duration: 60 min;Intervention duration: 12 weeks;Exercise frequency: 3 days a week;Exercise intensity: Unclear;Control: Usual care.SRSFu et al. 2025China56/6.06/44.64%(1)Exercise modality: Structured martial arts games;Session duration: 60 min;Intervention duration: 24 weeks;Exercise frequency: 3 days a week;Exercise intensity: Unclear;Control: Traditional rehabilitation training.ATEC, ABCShanker et al. 2023India43/9.47/18.60%(3)Exercise modality: Yoga;Session duration: 45 min;Intervention duration: 12 weeks;Exercise frequency: 5 days a week;Exercise intensity: Unclear;Control: Traditional rehabilitation training.ABC, SRSFeng et al. 2017China40/6.31/Unclear(1)Exercise modality: Aquatic Therapy (Halliwick Technology);Session duration: 90 min;Intervention duration: 4 weeks;Exercise frequency: 5 days a week;Exercise intensity: Unclear;Control: Routine rehabilitation.ABCNote: ADOS for Autism Diagnosis Observation Scale; GARS-2 for Gilliam Autism Rating Scale-Second Edition; SRS for Social Responsiveness Scale; SP for Sensory Profile; ATEC for Autism Treatment Evaluation Checklist; ABC for Aberrant Behavior Checklist; ASSS for Autism Social Skill Scale; RBS for Repetitive Behavior Scale; SSIS for Social Skills Improvement System; (1) for No significant baseline differences; (2) for Partially insufficient baseline reporting; (3) for Significant baseline differences present Characteristics of included studies and intervention features Exercise modality: Visuomotor / motor; Session duration: 50 min; Intervention duration: 10 weeks/Two months follow up; Exercise frequency: 3 days a week; Exercise intensity: Unclear; Control: Visual training/ Routine care. Exercise modality: Kata techniques training; Session duration: 30–90 min; Intervention duration: 14 weeks/One month follow up; Exercise frequency: 4 days a week; Exercise intensity: Unclear; Control: No-exercise control. Exercise modality: Kata techniques training; Session duration: 30–90 min; Intervention duration: 14 weeks/One month follow up; Exercise frequency: 4 days a week; Exercise intensity: Unclear; Control: No-exercise control. Exercise modality: Therapeutic horseback riding; Session duration: 60 min; Intervention duration: 12 weeks; Exercise frequency: 1 day a week; Exercise intensity: Unclear; Control: Wait-list control. Exercise modality: Nei Yang Gong; Session duration: 60 min; Intervention duration: 4 weeks; Exercise frequency: 2 days a week; Exercise intensity: Unclear; Control: Progressive muscle relaxation. Exercise modality: Therapeutic horseback riding; Session duration: 45 min; Intervention duration: 10 weeks; Exercise frequency: 1 day a week; Exercise intensity: Unclear; Control: Barn activity control intervention. Exercise modality: Combined physical training; Session duration: 60–80 min; Intervention duration: 8 weeks; Exercise frequency: 3 days a week; Exercise intensity: Unclear; Control: Did not perform any regular physical training. Exercise modality: Yoga; Session duration: 45–50 min; Intervention duration: 8 weeks/One month follow up; Exercise frequency: 3 days a week; Exercise intensity: Unclear; Control: Daily program activities. Exercise modality: AllPlay Dance; Session duration: 60 min; Intervention duration: 8 weeks; Exercise frequency: 1 day a week; Exercise intensity: Unclear; Control: Wait-list control. Exercise modality: Technical aquatic training/Game-based aquatic training; Session duration: 50 min; Intervention duration: 8 weeks; Exercise frequency: 2 days a week; Exercise intensity: Unclear; Control: Daily program activities. Exercise modality: Kata techniques training; Session duration: 30–90 min; Intervention duration: 14 weeks/One month follow up; Exercise frequency: 4 days a week; Exercise intensity: Unclear; Control: No-exercise control. Exercise modality: Gross motor intervention; Session duration: 60 min; Intervention duration: 6 weeks; Exercise frequency: 4 days a week; Exercise intensity: Moderate; Control: Daily program activities. Exercise modality: Mini-basketball Training; Session duration: 40 min; Intervention duration: 12 weeks; Exercise frequency: 5 days a week; Exercise intensity: Moderate (128–148 beats/minutes); Control: Routine rehabilitation. Exercise modality: Group sports game; Session duration: 60 min; Intervention duration: 12 weeks; Exercise frequency: 5 days a week; Exercise intensity: Unclear; Control: Routine rehabilitation. Exercise modality: Sports games combined with music therapy; Session duration: 30 min; Intervention duration: 12 weeks; Exercise frequency: 5 days a week; Exercise intensity: Unclear; Control: Music therapy. Exercise modality: Mixed martial arts training; Session duration: 45 min; Intervention duration: 13 weeks; Exercise frequency: 2 days a week; Exercise intensity: Unclear; Control: Wait-list control. Exercise modality: Basketball/Karate; Session duration: 60 min; Intervention duration: 12 weeks; Exercise frequency: 3 days a week; Exercise intensity: Unclear; Control: Usual care. Exercise modality: Structured martial arts games; Session duration: 60 min; Intervention duration: 24 weeks; Exercise frequency: 3 days a week; Exercise intensity: Unclear; Control: Traditional rehabilitation training. Exercise modality: Yoga; Session duration: 45 min; Intervention duration: 12 weeks; Exercise frequency: 5 days a week; Exercise intensity: Unclear; Control: Traditional rehabilitation training. Exercise modality: Aquatic Therapy (Halliwick Technology); Session duration: 90 min; Intervention duration: 4 weeks; Exercise frequency: 5 days a week; Exercise intensity: Unclear; Control: Routine rehabilitation. Note: ADOS for Autism Diagnosis Observation Scale; GARS-2 for Gilliam Autism Rating Scale-Second Edition; SRS for Social Responsiveness Scale; SP for Sensory Profile; ATEC for Autism Treatment Evaluation Checklist; ABC for Aberrant Behavior Checklist; ASSS for Autism Social Skill Scale; RBS for Repetitive Behavior Scale; SSIS for Social Skills Improvement System; (1) for No significant baseline differences; (2) for Partially insufficient baseline reporting; (3) for Significant baseline differences present Risk of bias of included studies According to the RoB 2.0 assessment, 12 studies were judged as having a low risk of bias in the domain of randomization process. In the domain of deviations from intended interventions, 2 studies were rated as high risk of bias, 1 study as low risk, and 17 studies as having some concerns. For missing outcome data, 10 studies were judged as having some concerns, while the remaining 10 studies were rated as low risk. In the domain of measurement of the outcome, 1 study was assessed as high risk of bias, 2 studies as low risk, and 17 studies as having some concerns. Regarding selection of the reported result, 19 studies were judged as low risk of bias and 1 study as having some concerns. Overall, 3 studies were classified as having a high risk of bias, while the remaining 17 studies were judged as having some concerns. The detailed risk-of-bias assessments are presented in Fig. 2. Risk of bias assessment using the RoB 2.0 tool Meta-analysis results For the intervention effects of exercise, a total of 20 studies contributing 85 effect sizes were included in the meta-analysis. The three-level random-effects model demonstrated a statistically significant overall effect, indicating that exercise significantly improved core ASD symptoms (g = -0.53, 95% CI -0.76 to -0.30, P < 0.001; 95% PI -1.35 to 0.29). For comparison, a conventional two-level random-effects model was also fitted and yielded a similar statistically significant pooled effect (g = -0.54, 95% CI -0.66 to -0.42, P < 0.001; 95% PI -0.66 to -0.42). The forest plot of the pooled effects is presented in Fig. 3. Intervention effects of exercise on core symptoms in children with ASD For follow-up outcomes, a total of five studies contributing eight effect sizes were included in the meta-analysis. The three-level random-effects model indicated that the overall maintenance effect of exercise on core ASD symptoms did not reach statistical significance (g = -0.46, 95% CI -0.97 to 0.06, P = 0.074; 95% PI -0.97 to 0.06). Similarly, the conventional two-level random-effects model yielded a comparable pooled effect estimate that also did not reach statistical significance (g = -0.46, 95% CI -0.97 to 0.06, P = 0.074; 95% PI -0.97 to 0.06). The forest plot of the maintenance effects is presented in Fig. 4. Maintenance effects of exercise interventions on core symptoms in children with ASD Sensitivity and influence analyses For the intervention effects of exercise, LOO sensitivity analyses showed that the pooled effect size remained stable after the exclusion of any single study, with estimates ranging approximately from -0.52 to -0.55. The overall conclusions and statistical significance were unchanged in all cases (all P < 0.001). In addition, heterogeneity indices showed minimal variation following study exclusion (τ² ≈ 0, I² close to 0), indicating good robustness of the findings. Influence diagnostics revealed that a small number of effect sizes exhibited relatively large studentized residuals (|rstudent| > 2); however, their Cook’s distance and DFFITS values indicated a limited influence on the overall pooled effect. Moreover, LOO analyses did not suggest that these effect sizes had a decisive impact on the overall conclusions. Therefore, no studies were excluded, and all eligible studies were retained in the primary analysis. Detailed results are presented in Supplementary Figures S1 and Table S2. For the maintenance effects, LOO sensitivity analyses similarly demonstrated stable pooled effect estimates following the exclusion of any individual study, with effect sizes ranging approximately from -0.41 to -0.53. The overall conclusions remained unchanged across all iterations, and the pooled maintenance effect remained statistically non-significant in each case. Heterogeneity indices again showed minimal variation after study exclusion (τ² ≈ 0, I² close to 0), supporting the robustness of the results. Influence diagnostics did not identify any outlying or influential effect sizes. Accordingly, no studies were excluded, and all studies were retained in the primary analysis. The corresponding results are shown in Supplementary Figures S2 and Table S3. In addition, given that two included studies contained effect sizes with baseline differences and another two studies provided insufficient information on baseline comparability, further sensitivity analyses were conducted. First, effect sizes with significant baseline differences were excluded and the pooled analyses were re-run. Subsequently, on this basis, studies with insufficient reporting of baseline comparability were further excluded, and the pooled analyses were repeated. These two-step sensitivity analyses were performed to examine the robustness of the findings. The results showed that, after excluding effect sizes with baseline differences, the pooled effect size changed only minimally (three-level random-effects model: g = -0.52, 95% CI: -0.73 to -0.31, P < 0.001; two-level random-effects model: g = -0.53, 95% CI: -0.65 to -0.40, P < 0.001). After further excluding studies with insufficient baseline comparability reporting, the pooled effect size still remained largely unchanged (three-level random-effects model: g = -0.55, 95% CI: -0.78 to -0.33, P < 0.001; two-level random-effects model: g = -0.58, 95% CI: -0.71 to -0.44, P < 0.001). In addition, sensitivity analyses excluding effect sizes derived from multi-component interventions were performed [35, 49]. The pooled effect size again changed only minimally (three-level random-effects model: g = -0.51, 95% CI: -0.76 to -0.26, P < 0.001; two-level random-effects model: g = -0.52, 95% CI: -0.65 to -0.39, P < 0.001). These findings further support the robustness of the main results. Heterogeneity analysis For the intervention effects of exercise, the Q test yielded Q(94) = 86.96 (P = 0.391), indicating that overall heterogeneity was not statistically significant. LRT further showed that the between-study variance component was not statistically significant (LRT = 0.00, P > 0.490), suggesting negligible heterogeneity between studies. In contrast, the within-study variance component was statistically significant (LRT = 18.78, P < 0.001), indicating substantial heterogeneity among effect sizes within the same studies. Regarding the proportion of total variance, Level 1 accounted for 67.19% of the total variability, Level 2 accounted for 32.81%, and Level 3 accounted for a negligible proportion of the total variability (approximately 0.00%). These findings indicate that heterogeneity in this meta-analysis was primarily driven by differences among multiple effect sizes within individual studies, rather than by variability between studies. Therefore, the application of a three-level meta-analytic model was justified, as the main source of heterogeneity resided at the within-study level rather than the between-study level. For the follow-up (maintenance) outcomes, the Q test yielded Q(7) = 1.20 (P = 0.991), indicating no statistically significant overall heterogeneity. Likelihood ratio tests showed that neither the between-study variance component (LRT = 0.00, P > 0.490) nor the within-study variance component (LRT = 0.00, P > 0.490) reached statistical significance. In terms of variance decomposition, Level 1 accounted for 100% of the total variability, whereas Level 2 and Level 3 accounted for approximately 1.30% and 1.81%, respectively. Moderator analyses Because only five studies provided follow-up data, moderator analyses were not conducted for maintenance effects. Therefore, moderator analyses were performed only for the intervention effects of exercise. The following variables were examined as potential moderators: core symptom domain, age, exercise modality, session duration, intervention duration, exercise frequency, and type of control condition. Detailed results are presented in Table 2. Table 2Moderator analysis of intervention effectsModerator AnalysesKg (95% CI) P Test of ModeratorsLevel 2Level 3Core symptomsF(1, 83) = 0.11,P = 0.73932.22%0.00%SCI53-0.51 (-0.77, -0.25)< 0.001RRB32-0.56 (-0.85, -0.27)< 0.001AgeF(1, 83) = 9.23,P = 0.00316.61%0.00%Preschool children26-0.96 (-1.28, -0.63)< 0.001School-aged children59-0.36 (-0.57, -0.15)0.001Exercise modalityF(3, 81) = 0.74,P = 0.53332.70%0.00%Traditional fitness/skill exercise48-0.61 (-0.92, -0.30)< 0.001Martial arts-based exercise15-0.64 (-1.08, -0.19)0.006Aquatic/sensory-based exercise10-0.51 (-1.12, 0.10)0.101Mind-body exercise12-0.12 (-0.73, 0.48)0.681Session durationF(1, 83) = 0.02,P = 0.87534.69%0.00%< 60 min33-0.55 (-0.92, -0.18)0.004≥ 60 min52-0.51 (-0.82, -0.20)0.002Exercise frequencyF(1, 83) = 3.40,P = 0.06931.07%0.00%< 3 sessions/week20-0.19 (-0.62, 0.24)0.378≥ 3 sessions/week65-0.66 (-0.92, -0.39)< 0.001Intervention durationF(2, 82) = 2.07,P = 0.13329.21%0.00%4–9 weeks25-0.22 (-0.61, 0.18)0.27910–12 weeks50-0.72 (-1.02, -0.42)< 0.001> 12 weeks10-0.50 (-1.04, 0.04)0.069Type of control conditionF(2, 82) = 4.80,P = 0.01120.63%0.00%Passive control45-0.33 (-0.59, -0.06)0.016Usual care control27-1.01 (-1.37, -0.65)< 0.001Active control13-0.42 (-0.84, -0.00)0.048Note: SCI for Social Communication and Interaction; RRB for Restricted and Repetitive Behaviors Moderator analysis of intervention effects F(1, 83) = 0.11, P = 0.739 F(1, 83) = 9.23, P = 0.003 F(3, 81) = 0.74, P = 0.533 F(1, 83) = 0.02, P = 0.875 F(1, 83) = 3.40, P = 0.069 F(2, 82) = 2.07, P = 0.133 F(2, 82) = 4.80, P = 0.011 Note: SCI for Social Communication and Interaction; RRB for Restricted and Repetitive Behaviors The analyses indicated that core symptom domain (F(1, 83) = 0.11, P = 0.739), exercise modality (F(3, 81) = 0.74, P = 0.533), session duration (F(1, 83) = 0.02, P = 0.875), intervention duration (F(2, 82) = 2.07, P = 0.133), and exercise frequency (F(1, 83) = 3.40, P = 0.069) were not significant moderators of the intervention effects. Age was identified as a significant moderator (F(1, 83) = 9.23, P = 0.003). Pairwise comparisons revealed that exercise interventions produced greater improvements in preschool children than in school-aged children (t = − 3.04, P = 0.003). In addition, the type of control condition was a significant moderator (F(2, 82) = 4.80, P = 0.011). Pairwise comparisons indicated that, compared with usual care control, both passive control (t = 3.21, P = 0.002) and active control (t = 2.47, P = 0.016) differed significantly in intervention effects. To further address the potential loss of information resulting from age categorization, age was additionally entered as a continuous moderator in a three-level meta-regression model. The results showed that mean age significantly moderated the intervention effects (F(1, 83) = 22.20, P < 0.001). Specifically, the regression coefficient for age was positive and statistically significant (β = 0.15, 95% CI: 0.09 to 0.21, P < 0.001), indicating that the pooled effect size became less negative with increasing age. Given that more negative Hedges’ g values indicate greater improvement in core ASD symptoms, this finding suggests that the beneficial effects of exercise tended to decrease as age increased (see Supplementary Table 2). Multiple moderator model A multiple moderator model was fitted using an active control condition and school-aged children as the reference categories. The overall moderating effect of the model was statistically significant (F(3, 81) = 3.86, P = 0.012), indicating that at least one moderator had a regression coefficient significantly different from zero. Detailed results are presented in Table 3. Table 3Results of the multiple moderator regression modelModeratorkβ95% CI P Intercept-0.86-1.47, -0.250.006AgePreschool children 26-0.13-0.79, 0.540.704Type of control conditionActive control 130.61-0.17, 1.390.126Passive control 450.54-0.10, 1.190.095 Results of the multiple moderator regression model Dose-response analysis Both nonlinear spline models and linear models were applied to examine the dose-response relationships between exercise exposure (total exercise volume and weekly exercise duration) and the intervention effects of exercise on core ASD symptoms. The corresponding results are illustrated in Figs. 5 and 6. Association between the nonlinear spline model and intervention effects Association between the linear model and intervention effects For total exercise volume, neither the nonlinear spline model (F(3, 81) = 1.93, P = 0.131) nor the linear model (F(1, 83) = 2.99, P = 0.087) showed a statistically significant association with intervention effects. Similarly, for weekly exercise duration, neither the nonlinear spline model (F(3, 81) = 0.58, P = 0.631) nor the linear model (F(1, 83) = 1.13, P = 0.290) demonstrated a statistically significant dose-response relationship with improvements in core ASD symptoms. Publication bias Publication bias was assessed using visual inspection of funnel plots and Egger’s regression test. For the intervention effects of exercise, the funnel plot showed a certain degree of asymmetry. After applying the trim-and-fill method to the left side of the funnel plot, the adjusted pooled effect size was g = -0.70 (95% CI -0.82 to -0.58). The magnitude of the effect size changed slightly, while the direction and statistical significance remained consistent. In addition, Egger’s regression test indicated no evidence of small-study effects (t = -1.41, P = 0.159) For the follow-up (maintenance) outcomes, the funnel plot appeared largely symmetrical, and Egger’s regression test also suggested no evidence of small-study effects (t = -0.14, P = 0.892). The corresponding funnel plots and trim-and-fill results are presented in Supplementary Figures S3 and S4. Certainty of evidence For the intervention effects of exercise, the certainty of evidence was evaluated using the GRADE approach. Regarding risk of bias, 17 included studies were judged as having some concerns and 3 studies as high risk of bias according to the RoB 2.0 assessment; therefore, the evidence was downgraded by one level. Although the funnel plot showed a certain degree of asymmetry in the domain of publication bias, the trim-and-fill analysis resulted in only a small change in the pooled effect size with a consistent direction, and Egger’s regression test was not statistically significant; thus, no downgrade was applied for publication bias. No downgrading was applied for inconsistency, indirectness, or imprecision. Consequently, the overall certainty of evidence for the intervention effects was rated as moderate. For the maintenance effects of exercise, only five studies were available, and none of the included studies were judged as having a low risk of bias according to the RoB 2.0 assessment. The total sample size was relatively small (n = 167), and the pooled effect estimate was g = -0.46 (95% CI -0.97 to 0.06), although no evidence of publication bias was detected. In the risk of bias domain, the evidence was downgraded by one level because all included studies were rated as having some concerns or high risk of bias. In the inconsistency domain, no downgrade was applied because the small number of studies (n = 5) limited the meaningful assessment of heterogeneity. In the imprecision domain, the evidence was downgraded by one level due to the small sample size and the wide 95% CI crossing the null value. No downgrade was applied for publication bias. As a result, the overall certainty of evidence for the maintenance effects was rated as low. Discussion This systematic review and three-level meta-analysis included 20 RCTs to examine the intervention effects, potential moderators, and dose–response relationships of exercise on core symptoms in children with ASD. However, only five of the included studies reported follow-up data after the completion of the intervention. The findings suggest that exercise was associated with improvement in core ASD symptoms, supported by moderate-certainty evidence; however, these findings should be interpreted with caution. In contrast, the maintenance effects of exercise were supported by low-certainty evidence and did not reach statistical significance. Intervention and maintenance effects of exercise on core ASD symptoms The present study found that exercise was associated with a statistically significant improvement in core symptoms in children with ASD, with an effect size of moderate magnitude (g = -0.53). However, this finding should be interpreted with appropriate caution. Although the pooled effect was statistically significant, the 95% PI crossed zero, indicating that the true effect may vary across settings and that some future studies may observe little or no benefit. In addition, most included studies were judged as having some concerns or high risk of bias, which further limits confidence in the strength and consistency of the observed effects. This finding is consistent with the overall conclusions of previous systematic reviews and meta-analyses. For example, Wang et al. synthesized 16 studies using a conventional meta-analytic approach and reported that exercise interventions were effective in improving core ASD symptoms, particularly deficits in social communication and interaction as well as restricted and repetitive behaviors [22]. Similarly, a network meta-analysis by Li et al. suggested that exercise can improve core symptoms in children with ASD, while different symptom domains may respond preferentially to different exercise components [23]. Distinct from previous studies, the present meta-analysis treated core ASD symptoms as a unified outcome, rather than pooling social communication deficits and repetitive behaviors separately. This approach aligns more closely with the DSM-5 conceptualization of ASD as a condition characterized by an integrated constellation of core symptoms, and it also reduces potential information loss arising from selective outcome inclusion or arbitrary aggregation of effect sizes. Notably, recent meta-analytic evidence has also shown that exercise significantly improves important functional outcomes in individuals with ASD, such as cognitive function and quality of life [14]. Although core symptoms were not the primary outcomes in that study, these findings indirectly support the beneficial role of exercise in improving overall functioning in ASD and suggest that improvements in cognitive and adaptive functioning may mediate or facilitate the effects of exercise on core symptoms. Although the pooled analyses showed very low between-study heterogeneity, this finding should also be interpreted cautiously. Given the diversity in exercise modalities, intervention protocols, and outcome measures, some degree of heterogeneity might have been expected. One possible explanation is that the present review included only RCTs and further restricted the sample to children aged 4–12 years, which may have reduced variability in study design and participant characteristics. In addition, although different instruments were used, they were all intended to assess core ASD symptoms, which may have increased conceptual comparability across studies. Importantly, the three-level model indicated that a greater proportion of variability was attributable to within-study differences among multiple effect sizes rather than to between-study heterogeneity. At the same time, the relatively small number of studies in some analyses may also have limited the statistical power to detect true between-study heterogeneity. Therefore, the low heterogeneity observed in the present study should not be interpreted as indicating complete homogeneity of intervention effects. Several mechanisms may underlie the beneficial effects of exercise on core ASD symptoms. First, exercise has been shown to promote neuroplasticity and improve brain network functioning, including enhanced functional connectivity between the prefrontal cortex and sensorimotor-related brain regions, which may support improvements in social information processing and behavioral regulation [55]. Second, exercise can enhance executive function and attentional control, which may help reduce impulsivity and repetitive behaviors while increasing behavioral flexibility and interaction quality in social contexts [7, 8]. In addition, exercise may indirectly alleviate restricted and repetitive behaviors and promote social participation by improving sensorimotor integration, modulating physiological arousal, and optimizing the allocation of attentional resources [22]. Despite the observed intervention effects, this study did not identify statistically significant maintenance effects of exercise on core ASD symptoms. Several factors may account for this finding. From a methodological perspective, the number of studies included in the follow-up analysis was limited. Only five RCTs, contributing a total of eight follow-up effect sizes, reported post-intervention follow-up data, which likely reduced statistical power to detect true maintenance effects. Furthermore, substantial variability in follow-up time points and assessment protocols may have weakened the ability to detect sustained effects. For example, follow-up durations varied across studies (e.g., one-month vs. two-month follow-ups [35]), and several studies included only a single follow-up assessment, limiting the ability to characterize the temporal trajectory of intervention effects [56]. Such heterogeneity in follow-up design may reduce comparability across studies and dilute pooled effect estimates. In addition, the absence of continued exercise stimulation after the intervention period may be a critical factor contributing to the lack of sustained effects. Unlike certain physical fitness or cognitive outcomes, core ASD symptoms, such as deficits in social communication and restricted and repetitive behaviors-are considered relatively developmentally stable and may require prolonged and consistent environmental input and behavioral shaping to achieve lasting change [56]. Once structured exercise interventions are discontinued, behavioral improvements may attenuate over time in the absence of continued support from family or school environments. Moreover, substantial individual heterogeneity among children with ASD-including baseline symptom severity, comorbid conditions, and individual responsiveness to exercise interventions-may further amplify variability during the follow-up phase, thereby reducing statistical significance at the group level. Therefore, the absence of significant maintenance effects in this study should be interpreted with caution. Rather than indicating that exercise lacks long-term benefits for core ASD symptoms, these findings more likely reflect limitations in the current evidence base regarding follow-up outcomes. Future high-quality RCTs should incorporate larger sample sizes, longer follow-up durations, and multiple follow-up assessments to more accurately evaluate the sustainability of exercise intervention effects on core ASD symptoms. Moderator effects: intervention characteristics With respect to intervention characteristics, exercise modality, session duration, intervention duration, and exercise frequency were examined as potential moderators. Among these variables, type of control condition emerged as a significant moderator, whereas exercise modality, session duration, intervention duration, and exercise frequency did not show statistically significant moderating effects at the omnibus level. Regarding exercise modality, martial arts-based exercise and traditional fitness/skill exercise were associated with significant improvements, whereas aquatic/sensory-based exercise and mind-body exercise did not demonstrate significant effects. This pattern may be explained, first, by limited statistical power due to insufficient evidence. Specifically, only three studies were available for each of the latter two exercise categories, and the small number of studies and effect sizes substantially reduces statistical power, making it more difficult to detect small-to-moderate effects even if they exist. Moreover, limited evidence increases the sensitivity of pooled estimates to individual study weights and design heterogeneity. Therefore, the observed non-significant findings are more likely to reflect insufficient or imprecise evidence rather than a true absence of effect. Second, previous systematic reviews have suggested that the benefits of aquatic exercise tend to be concentrated in outcomes such as motor skills, physical fitness, emotional functioning, and certain social behaviors [44, 57]. Behavioral improvements acquired in highly context-dependent aquatic environments may not readily transfer to land-based or real-world social interaction contexts, which could attenuate effects when core ASD symptoms-such as social communication deficits and restricted and repetitive behaviors-are assessed using standardized symptom scales. With respect to mind-body exercises (e.g., yoga and Tai Chi), existing reviews consistently indicate that studies in ASD populations remain limited in number, sample size, and methodological consistency, and current evidence is insufficient to support stable improvements in core symptoms [58]. Taken together, the lack of significant effects for these modalities is more plausibly attributable to limitations in the existing evidence base rather than a definitive lack of therapeutic potential. For session duration, both exercise sessions lasting < 60 min and those lasting ≥ 60 min were associated with improvements in core ASD symptoms. This finding differs somewhat from previous research. For example, Wang et al. reported that sessions lasting < 60 min and ≥ 90 min were effective, whereas sessions lasting 60–90 min were not [13]. Such discrepancies may partly arise from differences in duration categorization. The simplified dichotomization used in the present study (< 60 vs. ≥ 60 min) may have reduced the dilution of effects within intermediate duration ranges, thereby yielding more stable estimates. With respect to intervention duration, improvements were observed only for interventions lasting 10-12 weeks, whereas shorter (4-9 weeks) and longer (> 12 weeks) interventions did not show significant effects. This finding also differs from prior meta-analyses. For instance, Wang et al. categorized intervention duration as < 12 weeks versus ≥ 12 weeks and found no significant effects for shorter interventions [13]. The more refined temporal coding adopted in the present study may have revealed a potential “effective window.” Interventions that are too brief may be insufficient to induce stable behavioral and neural adaptations, whereas excessively long interventions without progressive overload or novelty may experience diminished effects due to reduced adherence, intervention fatigue, or habituation [59, 60]. Notably, a meta-analysis by Healy et al. found beneficial effects for interventions lasting < 10 weeks; however, their primary outcomes focused on physical fitness and functional measures rather than core ASD symptoms [61], suggesting that sensitivity to intervention duration may differ across outcome domains. Regarding exercise frequency, interventions delivered fewer than three days per week were not associated with significant improvements, whereas interventions conducted at least three days per week demonstrated beneficial effects. This finding is consistent with previous meta-analytic evidence [13]. From a theoretical perspective, higher exercise frequency may facilitate more consistent behavioral reinforcement and learning opportunities, promoting sustained engagement of executive function, self-regulation, and behavioral flexibility through repeated exposure to structured tasks and social cues [9]. In contrast, lower-frequency exercise may fail to generate sufficient cumulative behavioral stimulation, thereby limiting its impact on social communication deficits and restricted and repetitive behaviors [56]. Finally, type of control condition was identified as a significant moderator. Compared with usual care control, both passive control and active control conditions differed significantly in estimated intervention effects, suggesting that control conditions themselves can substantially influence effect size estimation. Previous systematic reviews in the ASD literature have emphasized that heterogeneity in control conditions is a major methodological source of inconsistency in exercise intervention effects, particularly when social communication and behavioral outcomes are the primary endpoints [56]. Therefore, future exercise intervention studies should not only carefully consider exercise modality and dosage but also clearly define and standardize control conditions to avoid biased estimation of intervention effects on core ASD symptoms. In addition, this finding may also be partly confounded by other study characteristics and should therefore be interpreted with caution. Moderator effects: other participant and outcome characteristics The present study also identified age as a significant moderator, with exercise interventions producing larger improvements in preschool children (g = -0.96) than in school-aged children (g = -0.36). This finding suggests that younger children may derive greater benefits from exercise-based interventions, potentially reflecting heightened neurodevelopmental plasticity during the preschool period. Evidence indicates that early childhood represents a critical window for the rapid development of social cognition, executive function, and sensorimotor integration, during which external environmental inputs-such as structured physical activity-are more likely to induce neural network reorganization and behavioral plasticity, thereby amplifying intervention effects [10, 62]. In contrast, core symptom patterns in school-aged children tend to be relatively more stable, and intervention-induced behavioral changes may therefore be more limited. In addition, stereotyped behaviors and social communication deficits in preschool children may not yet be fully entrenched. Early exercise interventions may facilitate improvements by enhancing self-regulation and behavioral flexibility, leading to more pronounced reductions in core ASD symptoms [63]. Previous studies have reported broadly similar patterns. For example, Wang et al. found that exercise improved core symptoms in both 3-6-year-old and 7-12-year-old children with ASD; however, they did not formally compare differences in effect magnitude between age groups [13]. Notably, when age was further examined as a continuous moderator, the results also showed a significant age-related trend, with older mean age associated with weaker intervention effects. This finding was consistent with the subgroup analysis indicating greater benefits in preschool children than in school-aged children, and it further supports the possibility that younger children may be more responsive to exercise-based interventions. One plausible explanation is that younger children may have greater neurodevelopmental plasticity and more malleable behavioral patterns, making them more responsive to structured environmental inputs, such as exercise-based interventions [60, 61]. In addition, the benefits of early intervention for reducing autistic symptoms and promoting adaptive development have been emphasized in previous studies [61]. This interpretation is also broadly consistent with previous meta-analytic evidence showing that exercise can improve core symptoms in both preschool- and school-aged children with ASD, although earlier work did not formally compare differences in effect magnitude across age groups [13]. Nevertheless, this finding should be interpreted cautiously because the meta-regression was based on study-level mean age rather than individual-level data, and developmental boundaries may vary across educational systems and studies. In contrast, core symptom domain was not identified as a significant moderator in the present study, indicating that exercise interventions were associated with improvements in both RRB and SCI. This finding is not entirely consistent with previous evidence. For instance, a meta-analysis by Yang et al. reported that exercise significantly improved RRB in children with ASD [64], whereas another meta-analysis by Huang et al. concluded that exercise did not significantly improve RRB [65]. Similar inconsistencies have been noted in earlier systematic reviews, with Bremer et al. highlighting that substantial heterogeneity in outcome selection, intervention protocols, and control conditions may contribute to divergent findings regarding RRB-related outcomes [56]. Several methodological factors may help explain these discrepancies. First, most previous meta-analyses treated SCI and RRB as independent outcomes and often selected only a single effect size per study, which may have resulted in the exclusion of relevant information and reduced the stability of effect estimates. Second, considerable variability exists in the measurement of RRB across studies, with both parent-reported rating scales and observational measures being used, further increasing heterogeneity and inconsistency across findings. In contrast, the present study employed a three-level meta-analytic model that explicitly accounted for dependency structures both within and between studies and incorporated multiple core symptom effect sizes reported within the same study. This approach allowed for more comprehensive and statistically efficient use of available data, potentially yielding more stable and generalizable estimates of exercise effects on both SCI and RRB. Dose-response relationship To our knowledge, this study is the first to systematically examine dose-response relationships between exercise and improvements in core ASD symptoms within a three-level meta-analytic framework, treating total exercise volume and weekly exercise duration as continuous variables. However, the findings indicated that neither total exercise volume nor weekly exercise duration emerged as significant moderators of intervention effects. Importantly, these results should not be interpreted as evidence that exercise dose is irrelevant, but rather as reflecting limitations in the current evidence base regarding dose-related variability. First, in previous exercise intervention studies in ASD populations, exercise dosage has typically been confined to relatively narrow and pragmatically feasible ranges (e.g., standard schedules in school or rehabilitation settings). Such restricted dose gradients may substantially reduce the ability to detect dose–response relationships, a pattern that has also been noted in prior systematic reviews and meta-analyses focusing on ASD [65, 66]. Second, total exercise volume and weekly exercise duration often covary with other intervention characteristics, including exercise modality, intensity, level of session structure, and the nature of control conditions. This collinearity and potential confounding may obscure the independent contribution of exercise dose to observed intervention effects.Moreover, for core ASD symptoms such as deficits in social communication and restricted and repetitive behaviors, improvements may be characterized by threshold or plateau effects. Once exercise interventions reach a minimal level sufficient to promote engagement and learning, further increases in duration may not yield proportional gains. Instead, improvements may depend more strongly on the quality of intervention content, the degree of structure, and the extent to which activities are embedded within meaningful social contexts [9]. Accordingly, the present findings suggest that exercise dose alone is unlikely to be the sole determinant of intervention effectiveness, and that structured program design and implementation quality may play a more critical role in improving core ASD symptoms. Strengths and limitations This study has several notable strengths. First, to our knowledge, it is the first meta-analysis to examine core ASD symptoms as the primary outcome and to systematically evaluate the effects of exercise interventions within a three-level meta-analytic framework. Beyond estimating intervention effects at post-intervention, this study further explored maintenance effects, moderators, and dose-response relationships, thereby providing a more comprehensive and nuanced evidence base for understanding how exercise may influence improvements in core ASD symptoms. Second, the inclusion criteria were strictly limited to RCTs, which helped reduce selection bias and confounding and enhanced the internal validity and reliability of the pooled estimates. Third, through systematic moderator analyses, this study examined potential sources of heterogeneity across multiple dimensions-including exercise modality, session duration, intervention duration, exercise frequency, age, and type of control condition-offering practical insights for the development and optimization of exercise-based intervention programs for children and adolescents with ASD. Nevertheless, several limitations should be considered when interpreting the findings. First, the literature search was restricted to studies published in Chinese and English, and studies published in other languages were not included. This restriction may have introduced language bias and could limit the generalizability of the results. Second, because this study focused on the overall level of core ASD symptoms, the included studies varied in outcome measures, assessment domains, and scoring methods. Although standardized effect sizes were used to synthesize results, differences in the sensitivity and construct emphasis of assessment instruments may have introduced measurement-related variability. Third, exercise intensity was unclear or insufficiently reported in most included studies, which prevented a more detailed examination of intensity as a potential moderator. This limitation restricts further interpretation regarding the optimal exercise prescription for children with ASD. Fourth, although the present meta-analysis used change-score-based effect sizes to reduce the potential influence of baseline between-group differences, baseline comparability was not reported with equal detail across all included studies. In addition, a small number of studies reported significant baseline differences in specific outcomes or subscales. Accordingly, the possibility that residual baseline imbalance may have affected some pooled estimates cannot be entirely excluded. Fifth, in the dose-response analyses, exercise dose was derived from nominal intervention parameters reported in the original studies (e.g., planned session duration, frequency, and intervention length). In addition, the range of exercise doses in the included ASD studies was relatively restricted, which may have limited the statistical power to detect dose-response relationships. In practice, children with ASD may not fully achieve the prescribed exercise dose due to factors such as attentional difficulties, adherence challenges, or symptom severity. Discrepancies between nominal dose and actual participation dose may have further attenuated the ability to detect dose-response relationships. Sixth, the number of studies included in the follow-up analyses was relatively small, and follow-up time points and assessment protocols varied across studies. These factors may have limited the statistical power to adequately evaluate the long-term maintenance effects of exercise interventions. Future directions Future research should prioritize high-quality RCTs with larger sample sizes and longer follow-up periods to more accurately assess the sustainability of exercise intervention effects on core ASD symptoms. In parallel, more detailed intervention reporting and objective monitoring methods are needed to quantify actual exercise participation dose, thereby improving the precision of dose–response analyses. In addition, future studies should further examine the moderating roles of individual characteristics, such as age, baseline symptom severity, and cognitive functioning, to support the development of more personalized exercise intervention strategies. Finally, integrating neuroimaging, physiological markers, and behavioral assessments to elucidate the underlying mechanisms through which exercise improves core ASD symptoms-and validating these interventions in school and community settings-will help enhance both the theoretical significance and real-world applicability of the evidence. Conclusions Based on the current evidence, exercise interventions may be associated with improvement in core symptoms in children with ASD, and these effects appear to be moderated by age and the type of control condition. However, the findings should be interpreted cautiously in light of the prediction interval crossing zero, the methodological limitations of the included studies, and the limited evidence regarding maintenance effects. No significant associations were observed between intervention effects and total or weekly exercise duration. Overall, exercise may be considered a promising supportive approach within the broader range of interventions for ASD, rather than a stand-alone or definitively established strategy. Its long-term sustainability and optimal intervention parameters still require confirmation through high-quality research. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author contributions QY Li contributed to the conceptualization of the study, literature search, data extraction, statistical analysis, interpretation of the results, and drafting of the manuscript. YM He, B Li, WQ Zhang, YY Wang, and HW Guan contributed to data extraction, data checking, statistical analysis, preparation of figures, and manuscript revision. ZFY Tuerdi, as the corresponding author and supervisor of the study, led the overall research design, established the methodological framework, supervised the systematic review and meta-analysis process, provided guidance on the PROSPERO-registered protocol, verified key decisions in study selection, data extraction, risk-of-bias assessment, and evidence grading, interpreted the findings, and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript. Funding This work was supported by A Study on the Historical Evolution and Contemporary Construction of Ethnic Minority Sports Culture in Xinjiang over the Past Century (22BTY106). Data availability The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This article is a systematic review and meta-analysis and does not involve any new studies with human participants performed by the authors. All included studies were conducted in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.