Which Parts of Cognition Actually Improve?
This chapter reads the exercise evidence at the resolution the title requires: not “does cognition improve?” in the abstract, but which cognitive abilities changed, by how much, and with how much consistency. The central subject remains repeated exercise in adults without diagnosed major cognitive impairment. The practical value of separating the outcomes is that a positive broad score, a small executive-function effect, and a mixed memory result can all belong to the same evidence base without being forced into one verdict.
Global cognition improves, but a broad score is not the whole mind
The broad sedentary-adult review found that exercise improved global cognition relative to control across nine studies. The pooled standardized mean difference was 0.57, with a 95% confidence interval from 0.11 to 1.03. The interval stays above zero, so the average difference favored exercise in the included sample. Heterogeneity was relatively low for this analysis: I² was 34.7%, meaning that the estimates varied, but not to the same degree seen in some other outcomes.
The resistance-focused review found a larger pooled estimate for global cognition: Hedges’ g was 0.76, with a 95% confidence interval from 0.32 to 1.21. But its heterogeneity was substantial, with I² = 68%. These two results point in the same broad direction while illustrating why effect sizes should not be treated as permanent properties of an exercise category. The estimates came from different reviews, populations, protocols, outcome measures, and comparison conditions. The resistance estimate’s greater variability also means that its average is a less complete description of what individual studies found.
A standardized mean difference is useful when studies use different tests or scales. It expresses the separation between groups in standard-deviation units rather than in raw points on one particular test. That makes unlike measures more comparable, but it removes the intuitive meaning of the original scale. A value of 0.57 does not mean an additional 0.57 IQ points, and it does not tell a particular reader how many minutes faster, how many items more accurately, or how much more independent they will become.
There is a second limitation. Global cognition is usually a composite or broad screening estimate built from several abilities. A composite can be more stable than any one task, but it can also smooth over a split result. Imagine an intervention that improves executive control and processing speed while leaving episodic memory unchanged. The overall score might rise, yet the composite would not tell us which part moved unless the component outcomes were examined separately. For that reason, the positive global-cognition finding is important evidence of broad average improvement, not a complete answer to the question of what changed.
Executive function is a clearer but conditional signal
Executive function is one of the more consistent domains in the two reviews. In the broad analysis, exercise versus control produced a standardized mean difference of 0.27, with a 95% confidence interval from 0.08 to 0.47. The estimate was statistically positive, but heterogeneity was moderate at I² = 48.4%. In the resistance-focused review, executive function also favored resistance exercise: Hedges’ g was 0.29, with a 95% confidence interval from 0.14 to 0.44 and I² = 50%.
These results concern abilities such as cognitive flexibility, inhibitory control, and problem solving. They are not a single “executive score” with one universal meaning. A task may ask a participant to suppress a prepotent response, switch rules, complete a trail under changing demands, or solve a problem that requires monitoring several steps. The common thread is control of goal-directed behavior, not memorization of a particular fact.
The broad review also found an age-related difference in its exploratory subgroup analysis. Studies whose participants had a mean age below 68 showed an executive-function effect of 0.50, with a 95% confidence interval from 0.19 to 0.80. Studies with a mean age of at least 68 showed an effect of 0.07, with a 95% confidence interval from −0.08 to 0.22. The interaction between the subgroups was statistically significant. That is a potentially important clue, but it is not a rule that exercise helps executive function below 68 and stops helping at 68.
The cutoff was the median age of the included studies, not a biological threshold. The younger group also covered a wide span of life stages, and the number of studies was limited. Differences between subgroups could reflect baseline cognition, health, protocol, measurement, or other characteristics that traveled with age. The finding supports further study of age as a possible moderator; it does not justify an individualized prediction from age alone.
The resistance result adds a useful counterweight to the age finding. Even without claiming that resistance exercise is best, its pooled executive-function effect suggests that the mode may support control processes in healthy adults. The mechanism chapter described possible physiological and attentional routes, but the domain result is what establishes the observed pattern. The proper conclusion is modest: executive function is a credible target for repeated exercise, while the size and applicability of the benefit still depend on who was studied and how the outcome was measured.
Resistance exercise sharpens the picture for working memory and processing speed
The resistance-focused review separates two abilities that are often hidden inside the broad word “cognition.” Working memory is the capacity to hold, update, and manipulate task-relevant information while an activity continues. Representative measures included Digit Span and Letter–Number Sequencing. Processing speed concerns the rapid completion of relatively simple operations, with measures that included reaction time and validated psychomotor-speed tests.
Resistance exercise produced a small but statistically significant pooled benefit for working memory: g = 0.37, with a 95% confidence interval from 0.19 to 0.55. Heterogeneity was low, at I² = 12%. It also produced a small positive effect for processing speed: g = 0.27, with a 95% confidence interval from 0.15 to 0.39 and I² = 7%. Within this review, those low heterogeneity estimates make the pattern more coherent than the episodic-memory result.
That coherence should not be inflated into a claim of unique superiority. The resistance review did not compare every resistance protocol directly with every aerobic or multicomponent protocol. A pooled effect against the included controls tells us that resistance exercise was associated with better performance on these outcomes than those comparison conditions. It does not prove that resistance exercise would beat aerobic exercise in a head-to-head trial, or that the effect would appear in every form of strength training.
The measurements also set the boundary of the conclusion. A faster reaction time may indicate more efficient processing under the test’s conditions, but it does not automatically mean faster reasoning in every setting. Better Digit Span performance suggests improvement in a task involving temporary information maintenance, but it does not demonstrate better long-term learning, judgment, or general intelligence. The result is valuable precisely because it is specific: resistance exercise appears promising for certain control and processing abilities, not because it carries a general label of mental enhancement.
Episodic memory is the clearest limit on a blanket claim
Memory is where the positive story becomes less stable. In the broad sedentary-adult review, the pooled effect for memory favored exercise but did not reach conventional statistical significance: SMD = 0.33, with a 95% confidence interval from −0.02 to 0.68. The interval crosses zero, so the data remain compatible with a small benefit, no average difference, or a result that would vary substantially across comparable studies. Heterogeneity was substantial at I² = 72.6%.
The resistance-focused review was even less decisive for episodic-memory retrieval. Its pooled estimate was g = −0.19, with a 95% confidence interval from −0.73 to 0.34 and I² = 83%. The interval includes both a possible disadvantage and a possible benefit, and the high heterogeneity indicates considerable variation between study estimates. This is not evidence that resistance exercise damages memory, nor is it proof that exercise has no memory value. It is evidence that the included studies do not provide a stable, consistently positive pooled result for this outcome.
The two memory estimates should not be combined as though they were measurements of one identical intervention. The first review pooled aerobic and multicomponent comparisons against control because its memory network was disconnected; it could not make a connected network comparison between the relevant exercise types. The resistance review focused on a different population and intervention family. Memory tasks also vary in encoding demands, delay length, retrieval cues, and scoring. Those differences can create genuine variation in what an outcome is sensitive to.
The broad review’s sensitivity analysis adds another reason to be careful. When the largest LIFE trial was omitted, heterogeneity fell and the pooled effect became slightly larger and statistically significant. But the trial met the prespecified eligibility criteria, so it remained in the primary analysis. A sensitivity result can show that one study influences the overall estimate; it does not authorize replacing the primary estimate with the more favorable alternative. The responsible conclusion is that memory may improve in some settings, but the current evidence does not establish a dependable average benefit for episodic retrieval.
How to read a result without overclaiming
The most useful reading method keeps four questions together. Direction asks whether the average estimate favors exercise or control. Magnitude asks how large the standardized difference is, while remembering that the number is not a personal unit of improvement. The confidence interval shows how precisely the studies locate the plausible average; an interval crossing zero means that a positive result is not firmly separated from no average difference. Heterogeneity asks how much the estimates differ across studies. High heterogeneity does not automatically invalidate a pooled result, but it warns that the average may not describe all protocols or participants equally well.
Statistical significance is only one part of that reading. A narrow, small effect can be statistically positive without being important to every person. A wider interval can contain a potentially meaningful benefit while leaving the average uncertain. A low I² value can make a pattern look coherent within the included studies, but it does not solve problems of measurement validity or applicability outside the studied population. Conversely, a high I² value does not prove that no intervention works; it signals that something differs across studies and needs explanation.
For example, the same pooled average could arise from several studies that all show small benefits, or from a mixture of strongly positive, null, and negative estimates. Those patterns have different implications even if their mathematical average is similar. The first suggests a more consistent response; the second suggests that participant characteristics, intervention details, control conditions, or measurement choices may determine who benefits. Heterogeneity is therefore not just a statistical penalty attached to a result. It is a clue that the broad label “exercise” may be hiding several distinct interventions and that the next question should be about moderators and applicability.
Certainty adds another layer. The broad review judged direct comparisons for global cognition and executive function to be of moderate quality, while most indirect and mixed comparisons were rated very low certainty because of study limitations, imprecision, and publication-bias concerns. The resistance review had a more favorable risk-of-bias profile overall, but still reported varied controls, attrition, measurement differences, and substantial heterogeneity for some domains. The direction of several findings is therefore more secure than any precise ranking of exercise modes.
The resulting domain map is neither “exercise changes everything” nor “exercise does nothing.” Global cognition and executive function show positive average signals in both evidence sets. Resistance exercise adds relatively coherent findings for working memory and processing speed. Episodic memory remains mixed, with wide variation between studies. The next question is how these domain patterns should affect the comparison between aerobic, resistance, and multicomponent exercise—and whether a ranking can be useful without being mistaken for a universal winner.