npj | complexity Article

https://doi.org/10.1038/s44260-025-00039-x

A first-principles mathematical model

integrates the disparate timescales of
human learning
Check for updates
1,2,6 3,6 4,5,6
Mingzhen Lu , Tyler Marghetis & Vicky Chuqiao Yang

Lifelong learning occurs on timescales ranging from moments to decades. People can lose
themselves in a new skill, practice for hours until exhausted, and pursue mastery intermittently over
decades. A full understanding of learning requires an account that integrates these timescales. Here, in
1234567890():,;
1234567890():,;

response to calls for more formal theory in the psychological sciences, we present a parsimonious
mathematical model that unifies the nested timescales of learning. Our model recovers well-
established patterns of skill acquisition, and explains how these patterns can emerge from short-
timescale dynamics of motivation, fatigue, and effort. Conversely, the model explains how patterns in
these short-timescale dynamics are shaped by longer-term dynamics of skill selection, mastery, and
abandonment. We use this model to explore the theoretical benefits and pitfalls of a variety of training
regimes. Our model connects disparate timescales—and the subdisciplines that typically study each
timescale in isolation—to offer a unified, multiscale account of skill acquisition.

Learning is a through-line of human life, from children at play to Picasso at his formal theory building in the psychological sciences4–6. Unlike statistical
peak. The process of learning occurs on multiple timescales1–3. Over the long models, which aim to describe empirical data, or cognitive process models,
timecourse of mastering a skill, learning may appear continuous and gradual which aim to capture in detail the internal mental processes that give rise to
as one progresses from novice to expert. But this slow process consists of particular psychological outcomes, our model offers a more foundational
countless shorter, discrete periods of engagement, lasting from mere account of common patterns found across superficially dissimilar domains
moments to entire days. An individual may start working but get bored. They of learning. This approach to theory building has been influential in fields
may rest. They may eventually master the skill, or give up, or switch focus to such as ecology7 and statistical physics8. Formal models such as the one
another skill entirely. In short, learning occurs on multiple nested timescales, introduced here are opportunities to specify the minimal components that
from the moment-to-moment dynamics of motivation and fatigue, to briefer suffice to give rise to qualitative patterns, thus advancing our theoretical
episodes of sustained effort, to the slow crawl toward mastery. understanding of the origin of those patterns. Here, we offer such an account
A full understanding of learning requires an account that integrates of multiscale lifelong learning.
these timescales3. Existing accounts, however, have typically focused on single Below, we briefly review the longer and shorter timescales of learning.
timescales. Here, we present a minimal mathematical model that unifies the We then introduce our formal model, which unifies these timescales.
nested timescales of lifelong learning, offering a theoretical account of well- An often-repeated cliché holds that mastery requires “10,000 h”—that
established patterns at both short- and long-timescales. The model captures is, two years of constant, daily effort, breaking only to sleep at night9,10. While
the relations between the dynamics of motivation, fatigue, and sustained overly simplifying, this saying captures the fact that real mastery requires
engagement on the shorter timescale of individual periods of work (e.g., sustained engagement over long periods9–13.
minutes or hours), as well as well-established, classic patterns of learning and The dynamics of learning on this timescale are typically described
performance on the longer timescale of mastery (e.g., days or years). using so-called “learning curves” or “progress curves”12,14–16. These represent
Our goal is to offer a formal explanation of why similar qualitative the gradual increase in skill over time, where skill is typically measured at the
patterns in skill learning have been found to occur in a wide range of level of an entire session of work or practice: the final score in a game, the
domains, from juggling to chess. We are thus responding to the call for more best performance in some athletic skill within a training session, mean

1
Santa Fe institute, 1399 Hyde Park Road, Santa Fe, NM, 87501, USA. 2Department of Environmental Studies, New York University, 285 Mercer Street, New York,
NY, 10003, USA. 3University of California, Merced, 5200 Lake Rd, Merced, CA, 95343, USA. 4MIT Sloan School of Management, Massachusetts Institute of
Technology, Cambridge, MA, 02142, USA. 5Institute for Data, Systems, and Society, Massachusetts Institute of Technology, Cambridge, MA, 02142, USA. 6These
authors contributed equally: Mingzhen Lu, Tyler Marghetis, Vicky Chuqiao Yang. e-mail: [email protected]; [email protected]; [email protected]

npj Complexity | (2025)2:15 1

https://doi.org/10.1038/s44260-025-00039-x Article

it is now the standard approach at the highest levels of the sport23. Thus, on
the slow timescale of skill mastery, individuals typically exhibit learning with
diminishing returns, and may exhibit repeated plateaus, followed by brief
dips in performance and increased learning. This general pattern has been
documented in a striking range of domains, from juggling to mental
arithmetic, from sports to video games23.
The slow process of skill mastery, however, is built out of many brief
sessions of practice and learning25,26. People choose to start working, and
then persist in that effort until they give up or require rest—and then, having
rested, they may choose to start once more. When people start and stop is
often a function of their motivation27 and fatigue28 (Exceptions include cases
where people are forced to work by some external entity, such as a teacher,
coach, or employer). While the development of skill and performance are
typically measured at the level of entire work-sessions, people are never-
theless learning within each session. On this fast timescale of moment-to-
moment engagement, motivation and fatigue may fluctuate, skills may
creep ever upward, and accomplishments may accumulate slowly. Long-
term mastery, therefore, emerges from the the dynamics of short-timescale
engagement and learning.
A full understanding of life-long learning and sustained engagement
requires an understanding of both long and short timescales3. The precise
timescale on which “long timescale” learning occurs will depend on the
particular task and domain. Mastering a new video game, for instance, may
Fig. 1 | Classic patterns in the dynamics of learning, performance, and sustained require hours or days, while becoming a chess master can take decades.
engagement. Our mathematical model aims to recover these qualitative patterns. From our perspective, these are both “long timescale” processes, since they
A Skill development (solid red) typically exhibits diminishing returns on practice as involve changes that are orders of magnitude slower than moment-to-
one approaches mastery of a task. Over time, one can switch to more challenging moment fluctuations in motivation and fatigue. At present, however, we
tasks (dashed blue line), thus creating opportunities for continued skill development. lack a unifying framework, one that situates known patterns on the short
Zoomed inset shows a typical learning curve with diminishing returns over time. timescale of single work sessions within the context of well-established
B Performance typically increases with practice but can reach plateaus. When one patterns of long-timescale skill development.
switches to a more challenging task, there can be a momentary dip in performance This is a problem because changes on one timescale shape changes on
before performance outstrips the previous plateau. C For people who are motivated the other29. Lifelong learning emerges from the moment-to-moment
by skill development, motivation typically peaks when the task is sufficiently difficult dynamics of motivation, fatigue, skill development, and performance30. The
to maximize learning. As the gap shrinks between skill and task difficulty, motivation
decision to work or rest, meanwhile, is a function of fatigue and motivation,
typically decreases. Switching to a more challenging task can increase motivation.
but motivation is shaped by long-timescale changes in skill and in the task
D Average work and rest period lengths. When individuals motivated by skill
development encounter a task that is sufficiently but not overly challenging, they
being learned31. The dynamics of each timescale only make sense in light of
typically work for long periods (orange), and require comparatively little rest (grey). the other.
Over time, as the gap shrinks between skill and task difficulty, they may work for Encouraged by successful models of nested timescales in other biolo-
shorter and shorter periods as they get increasingly bored. Note that these dynamics gical and social systems32–35, here we develop a first-principles mathematical
reflect endogenous behavior of the model without external forcings. model that unifies the nested timescales of learning. By bridging between
levels—the long timescale of mastery, and the short timescale of effort and
engagement—the model connects phenomena that have traditionally been
typing speed within a testing period, etc. In domains as varied as music, investigated by distinct research traditions. To be clear, we are not offering a
chess, memory, typing, sport, and juggling, progress curves typically show statistical model of empirical data from any particular study of skill acqui-
diminishing returns, with the greatest rate of learning when skill is far from sition. Moreover, any minimal model of a phenomenon as complex as
mastery, and a gradually diminishing rate of learning as mastery is human lifelong learning will only be able to account for major qualitative
approached (Fig. 1A, inset)12,14–17. Empirical progress curves can be descri- patterns. And outcomes that reflect factors that lie outside the model—such
bed using power or exponential functions, although there is debate about the as exogenous pressures to work or rest, whether driven by diurnal cycles or
best-fitting functional form18–22. workplace supervisors—will necessarily lie outside the model’s scope.
Indeed, more recent work has found that, at the individual level, Rather, our goal is to offer a task-agnostic account that explains, from first-
learning often improves in distinct stages (Fig. 1A), reflected in plateaus in principles, why similar patterns have been found to recur across domains,
the learning curve18,23,24. These plateaus are periods where the learner has from chess to juggling. Our contribution is thus theoretical, not empirical.
mastered a particular task and has not yet switched their focus to a more In what follows, we introduce the formal model. We show that the
challenging one. The learner may eventually pursue a more challenging task, model reproduces many classic, well-established patterns in skill acquisi-
which might result in worse performance at first but ultimately better tion, on multiple timescales. This first-principles model thus offers a theo-
performance as the new, more challenging task is mastered (Fig. 1B)23. In retical account of why similar patterns recur across domains. We then
some instances of this general scenario, the simpler “task” is a sub-optimal explore the model’s behavior in the context of different training regimes
approach to a problem to which the more difficult “tasks” are better but (e.g., continuously increasing task difficulty vs. discrete jumps in difficulty)
more challenging solutions. An example of this is when athletes in the sport and for individuals who differ in their motivation.
of high jump start with easier but less effective ways of jumping over the bar
(e.g., jumping forwards) but eventually transition to the more challenging The mathematical model
but ultimately more effective “Fosbury Flop” (i.e., jumping backwards). The We present a first-principles model that integrates the multiple timescales of
Fosbury Flop was originally derided as an overly difficult approach to the learning, from moment-to-moment engagement to life-long growth and
problem of jumping over the bar, and it is indeed more challenging at first skill acquisition. We take a dynamical systems approach, where our model
than jumping face first, but it is ultimately a more effective solution; indeed, consists of a system of ordinary differential equations. The dynamical

npj Complexity | (2025)2:15 2

https://doi.org/10.1038/s44260-025-00039-x Article

Fig. 2 | Conceptual illustration of the modeling framework and key variables. (solid blue curve). C The decision to work (p = 1) or rest (p = 0) is a function of the
A Short timescale variables (motivation, M, and fatigue, F) determine whether an difference between motivation and fatigue (M − F). Starting to work requires suf-
individual decides to work, p. As an individual works, skill (S) improves as does task ficient motivation relative to fatigue (i.e., M − F > θ). Work stops when fatigue
execution (E). Task difficulty (T) can change over time or remain constant. overtakes motivation (i.e., net motivation, M − F ≤ 0). The additional effort required
B Illustration of the distinction between long- and short-timescale variables. Long to start working, θ, induces path-dependence in the decision to work or rest (i.e.,
timescale variables (e.g., skill) tend to change slowly (dashed green line). Short hysteresis). D Skill S and execution E change more slowly, and their dynamics
timescale variables (e.g., motivation) change more rapidly, sometimes oscillating depend on the relative task difficulty (i.e., T − S).

Table 1 | Summary of key model variables and parameters

Variable Meaning Range

Long timescale variables
S Individual’s skill level [0, ∞)
E Individual’s rate of execution on task [0, ∞)
T Task difficulty [0, ∞)
Short timescale variables
M Individual’s current motivation [0, ∞)
F Individual’s current fatigue [0, ∞)
p Whether individual is working (1) or resting (0) {0, 1}
Parameters
Parameter Meaning Range Typical value
w Whether individual is more driven by the accumulation of achievements (0) or skills (1) [0, 1] 0.5
c1 Weight of contribution from return to baseline to motivation [0, ∞] 0.5
c2 Weight of contribution from learning and execution to motivation [0, ∞] 1.2
c3 Rate of fatigue increase while working [0, ∞] 0.1
c4 Rate of fatigue decrease while resting [0, ∞] 0.1
xm , ym Shape parameters for learning and execution function fS, fE [0, ∞] 1, 0.2
θmin Minium activation threshold to switch to working from resting [0, ∞] 0.005
ϵmin, k Parameters to determine the threshold for increasing task difficulty (ϵ) in stepwise scenario [0, ∞] 0.005, 0.01

systems approach has been highly effective at providing elegant explana- execution captures the fact that people often learn best and perform best at
tions of complex phenomena in human systems such as motor different times43. One performs best when executing tasks that are easy,
coordination36–38, cognitive development3,24, left-handedness in a minority relative to one’s skill level, while one develops skills when working on tasks
of the population39, and the polarization of political parties40,41. We divide that are more difficult (but not too much more difficult) than one’s current
the description of the dynamical system into two components: three state skill level44,45. We calculate an individual’s performance as the product of
variables that change on a long timescale, and another three that change on a their current skill and current execution, S(t)E(t), since performance is
short timescale. The mathematical model thus contains variables that jointly affected by skill level and rate of execution.
change on both short and long timescales, with the different timescales
connected by the individual’s decision to work or rest. Figure 2A illustrates Skill development and execution. In the model, skill improvement is a
the mathematical model’s conceptual framework and the relationships function of relative task difficulty, which is the difference between task dif-
between key variables. See Table 1 for a summary of key model parameters ficulty and one’s skill level (T(t) − S(t)). In other work, this is also referred to
on various time scales. as functional task difficulty44. Specifically, skill improves when the task is
harder than one’s current skill level (T(t) − S(t) > 0), but only to an extent. If
Long timescale a task is too easy or too difficult, the individual working on it does not
The three key long-timescale variables are S(t), the skill level of an individual develop skill44,45. Put mathematically, the dynamics of skill are given by:
at time t, T(t), the task difficulty, and E(t) (derived from S and T), the rate of
execution, or how much work an individual is getting done at time t. Within dSðtÞ
¼ pðtÞf S ðTðtÞ  SðtÞÞ; ð1Þ
psychology, rate of execution is sometimes referred to as efficiency42, though dt
here we use ‘rate of execution’ to avoid confusion with other uses of the term
‘efficiency’ in economics and engineering. The variables S, E, and T all take where p(t) is a boolean variable denoting whether the individual is working
continuous, non-negative values. The distinction between skill and at time t. If they are working, p(t) = 1, and if resting, p(t) = 0. Note that all

npj Complexity | (2025)2:15 3

https://doi.org/10.1038/s44260-025-00039-x Article

variables and parameters in this and the following equations are dimen- Task difficulty. Unlike skill (S(t)) and execution (E(t)), which are solely
sionless. In the Supplemental Information, we show how these dimen- governed by the internal dynamics of the dynamical system, our third
sionless equations are derived from dimensional ones. The function fS long-timescale variable, task difficulty (T(t)), can be treated as an external
represents how task difficulty, relative to the current skill level, affects the input variable. As a result, our modeling framework allows us to explore
rate of skill acquisition. For simplicity, we consider this function to take the how different ways of increasing task difficulty affect the dynamics of
shape of a triangle, peaking at value xm (where xm > 0), which we visualize in long-timescale learning and short-timescale engagement.
Fig. 2D. The function can be expressed in the following piecewise form: Specifically, we evaluated several different training regimes: task diffi-
culty remains unchanged (constant scenario), increases in a discrete fashion
8 once skill catches up with task difficulty (stepwise scenario), and increases in
>
>
0 if x<0;
>
< x=x a continuous fashion (continuous scenario).
m if 0 ≤ x ≤ xm ;
f S ðxÞ ¼ ym ð2Þ
> ðx þ 2xm Þ=xm
> if xm <x ≤ 2xm ; Short timescale
>
:
0 if x>2xm ; On the short timescale, three key variables that govern the individual’s
working behavior are motivation (M(t)), fatigue (F(t)), and work (p(t))46–48.
This piece-wise function peaks at height ym, denoting the maximum Motivation refers to one’s current willingness or desire to work on a task49.
learning rate. Note that this approach, unlike power-law or exponential For simple tasks that are primarily physical, this may be driven by levels of
learning curves, allows us to model the period of skill acquisition when the the neurotransmitter dopamine50; for more complex or conceptual tasks,
initial task is much too difficult for the learner, given their current skill. motivation intensity is likely a composite of different psychological and
During this period, learning is laborious and slow, with little skill acquisition physiological cues. Here, we remain agnostic about the specific physiological
occurring even with extended effort. Think of the young child who first underpinning of motivation. We model motivation, M(t), as a continuous
encounters chess before they are sufficiently mature to even remember all variable taking non-negative values.
the rules: repeated encounters with the game will produce little improve- Fatigue is a sense of cumulative exertion or tiredness51. For some simple
ment in their ability to play chess. This period is often ignored in empirical tasks that are primarily physical, the sense of fatigue may be driven by simple
work on skill acquisition, because participants and tasks are intentionally physiological processes, such as the accumulation of cerebral adenosine50;
matched so the task is approachable. However, in many real-world sce- for more complex or conceptual tasks, much like for motivation, the sense of
narios, an individual encounters a task that is too difficult to even start—or if fatigue is likely a composite of different psychological and physiological
starting is possible, then it is initially quite laborious. In our modeling cues. Here we also remain agnostic about the specific physiological under-
framework, this period corresponds to the right-hand side of the learning pinning of fatigue. We model fatigue, F(t), as a continuous variable taking
function where relative task difficulty is very high. non-negative values.
Second, our approach is agnostic to the exact functional form of the
learning curve. We show in the Supplementary Information that the Motivation. We assume the level of motivation is affected by two pro-
function fS, which governs the rate of learning as a function of the relative cesses. The first is that, in the absence of other influences, motivation will
task difficulty, can be modified to generate learning curves that follow, restore to a baseline52,53. In dimensionless equations, this baseline is
exactly, either an exponential or a power law curve. Here, we choose the represented as 1 (see Supplementary Information for details on the
simplest functional form that gives diminishing returns on effort: a piece- nondimensionalization process). The second is that motivation increases
wise linear relationship between relative task difficulty and rate of learning. when an individual is learning (Eq. (1)) or executing a task (Eq. (3))54,55.
This allows us to capture the qualitative features of empirical learning curves Mathematically, these dynamics are expressed as
with fixed tasks, while also allowing us to model more complex scenarios  
involving variable task difficulty (e.g., when task difficulty is simultaneously dMðtÞ dSðtÞ
¼ c1 ð1  MðtÞÞ þ c2 w þ ð1  wÞEðtÞSðtÞ ; ð5Þ
changing with time or adapting to skill). dt dt
Similarly, the rate of execution, a dimensionless quantity that indicates
the proportion of skill level brought to conduct the task, is given by, where the first term reflects the process of returning to baseline, and the
second term captures the increase in motivation due to learning and per-
formance. The parameters c1 and c2 are constants weighting the contribu-
EðtÞ ¼ pðtÞf E ðTðtÞ  SðtÞÞ; ð3Þ tion of these two processes. As introduced above, S(t) denotes skill and E(t)
denotes execution of an individual at time t. The relative influence of
The shape of the function fE is sketched in Fig. 2D, which is a positive learning and performance on motivation is determined by parameter w,
constant when relative task difficulty is negative (T(t) − S(t) < 0), and decays which takes values between 0 (only performance matters) and 1 (only
and eventually settles to 0 when relative task difficulty increases (T(t) − learning matters). Simplifying Eq. (5), we have
S(t) > 0). The function can be expressed in the following piecewise form:
dMðtÞ  
8 ¼ c1 ð1  MðtÞÞ þ c2 pðtÞ wf s ðTðtÞ  SðtÞÞ þ ð1  wÞSðtÞf E ðTðtÞ  SðtÞÞ :
dt
>
<1 if x < 0;
ð6Þ
f E ðxÞ ¼ ym ðxm  xÞ=xm if 0 ≤ x ≤ xm ; ð4Þ
>
: Fatigue. The level of fatigue is also affected by two processes. The first is
0 if x > xm :
that when one is working, the level of fatigue increases56. Here, we use the
simplest form of this process, a linear increase. The second is that when
Intuitively, this says that execution is at a maximum when the demands
one is resting, the level of fatigue recovers to baseline, zero. Here, we draw
of the task are fully met by one’s skill, but execution decreases linearly as the
on the literature on recovery (e.g.,ref.28), which has found that recovery
deficit between task difficulty and skill gets larger44.
tends to involve an exponential decay in fatigue. These dynamics of
An individual’s performance on a task reflects both their skill and
accumulating and decaying fatigue are mathematically integrated into
execution. Performance is highest when they are both highly skilled and
the following conditional function,
executing the task with maximum execution; it is lower for individuals with
lower skill, individuals who are struggling to execute the task, or both. We 
dFðtÞ c3 if pðtÞ ¼ 1;
thus measure an individual’s performance as the product of their skill and ¼ ð7Þ
their rate of execution, S(t)E(t). dt c4 FðtÞ if pðtÞ ¼ 0:

npj Complexity | (2025)2:15 4

https://doi.org/10.1038/s44260-025-00039-x Article

(A) (B) (C) (D) (E)

Fig. 3 | Constant task difficulty leads to plateauing skill and performance on the long both decline over time, consistent with the decreasing motivation. Note that in the
timescale (A, B), and reduced motivation and working periods on the short timescale absence of some additional mechanism for sustaining engagement, the individual
(C–E). A Skill (red solid line) approaches the task difficulty (blue dashed line), and initiates work but immediately abandons the task, as it is no longer sufficiently
eventually plateau. B Performance of an individual (S(t)E(t)), increases over time but motivating to sustain work for more than a moment. In the real world, an individual
eventually plateau. C, D Motivation and fatigue oscillate across recurring working at this point would likely abandon the task entirely or would require some other
sessions, where individual enter a work session when motivation is sufficiently mechanism to sustain their engagement, whether endogenous to the individual (e.g.,
greater than fatigue and exit the work session when motivation drops below fatigue. an explicit dedication to persisting despite the lack of reward) or exogenous (e.g., a
Peak motivation value of each work session diminishes as one’s skill and perfor- supervisor or coach who encourages sustained work).
mance plateau. E The durations of an individual’s working period and rest period

This differential equation formulation is equivalent to a linear increase empirical pattern of diminishing returns in skill learning when task diffi-
in time when working, and exponential decay when resting. Constants c3 culty is held constant12,14–16. In our model, this pattern emerges naturally
and c4 set the timescales of the changes. from short-timescale dynamics of motivation, fatigue, and engagement.
The model also predicts that long-timescale skill acquisition should
Work. An individual’s decision to work or rest depends on their moti- be accompanied by changes in short-timescale dynamics. As skill
vation and fatigue57,58. Recall that the variable p(t) indicates whether an improves, both peak motivation and peak fatigue within working ses-
individual is working (p(t) = 1) or resting (p(t) = 0). p(t) is determined by sions decline monotonically (Fig. 3C, D). Moreover, work sessions last
the difference between motivation and fatigue, M(t) − F(t). An indivi- longer—that is, engagement is more sustained—when the task difficulty
dual will switch to working from resting when they have sufficient is suitably higher than the skill level (Fig. 3E). Approaching the plateau,
motivation relative to their fatigue—that is, when M(t) − F(t) increases one becomes less and less motivated. Consequently, one tends to work
across a critical threshold, θ(t), which increases with the difference for shorter and shorter periods, suggesting an increasing degree of
between task difficulty and skill level, while limited to a minimum value, boredom.
θðtÞ ¼ maxðTðtÞ  SðtÞ; θmin Þ. The minimum θmin reflects the minimum The apparently static “plateau” in skill development at the macro-scale
motivational start-up cost of starting to work. Individual stops working is actually the product of an intricate dynamic balance of micro-scale pro-
when their fatigue is equal to or greater than their motivation—that is, cesses. Indeed, in the real-world, the decision to begin working reflects not
when M(t) − F(t) is smaller than or equal to zero. A sketch of this just the endogenous influences of motivation and fatigue captured by our
dynamic is shown in Fig. 2C. In discrete time, the function p(t) can be model, but also exogenous influences such as diurnal cycles (i.e., people are
expressed as not going to start working while sleeping). Moreover, individuals who are

insufficiently motivated by a task will likely require exogenous influences to
0 if pðtÞ ¼ 0 and MðtÞ  FðtÞ < θ; or pðtÞ ¼ 1 and MðtÞ  FðtÞ ≤ 0; persist in working—a workplace supervisor or a coach, for instance. Such
pðt þ ΔtÞ ¼
1 if pðtÞ ¼ 0 and MðtÞ  FðtÞ ≥ θ; or pðtÞ ¼ 1 and MðtÞ  FðtÞ > 0; exogenous influences are beyond the scope of our minimal model. In the
ð8Þ absence of such external pressures, individuals faced repeatedly with a task
that is so unrewarding that it can only sustain their engagement for a
where Δt is a small time interval. The activation threshold should be lower moment would likely abandon the task for something more challenging and
for individuals at higher skill levels relative to task difficulty. In other words, thus rewarding. The baseline scenario explored in Fig. 3 does not allow for
it is easier for an individual to start working on a task when they are this kind of adjustment; indeed, we assume for simplicity that individuals are
good at it. forced to repeatedly attempt a task that is no longer rewarding. In sub-
sequent sections, we relax this assumption to explore scenarios where the
Results task is dynamically adjusted or even abandoned entirely for a new, more
Diminishing return in skill and performance under constant task challenging task.
difficulty
We first analyze a baseline scenario where an individual experiences con- Sustained skill learning under stepwise increases in task
stant task difficulty over time, illustrated in Fig. 3A (hereafter constant difficulty
scenario). Real-world examples of this constant scenario are ubiquitous, at In real-world settings, individuals tend to engage in increasingly difficult
least within delimited time periods (e.g., lifting a certain weight, playing a tasks throughout the long-term learning process. To capture this, we eval-
card game against a fixed computer opponent). Note that this scenario is uated a scenario where task difficulty increases in discrete jumps whenever
analogous to the regime illustrated in the inset of Fig. 1A. We assume, an individual’s skill catches up with task difficulty, as shown in Fig. 4A
moreover, that the individual is motivated equally by learning and perfor- (hereafter stepwise scenario). Real-world situations corresponding to this
mance; in the last section of the Results, we weaken this assumption and situation include strength training with incrementally heavier weights,
explore the effects of individual differences in motivation. attempting a new level in a video game, or skipping a grade in school. We set
In this constant scenario, our model recovers the classic result that task difficulty to increase by a fixed increment of 0.8 whenever the gap
long-timescale skill development reaches a plateau of both skill and per- between an individual’s skill and the demands of the task (i.e., T − S) is
formance, as shown in Fig. 3A, B. The model thus generates the classic smaller than a threshold, ϵ.

npj Complexity | (2025)2:15 5

https://doi.org/10.1038/s44260-025-00039-x Article

Skill and task difficulty Performance Movaon Fague Work & rest period length
(A2) (A3) (A4)
(A5)
(A1)

(B1) (B2) (B3) (B4) (B5)

(C1) (C2) (C3) (C4) (C5)

Fig. 4 | Short- and long-timescale dynamics of three distinct learning regimes. slow pace. The learner adapts to increasing task difficulty by improving skill.
A Model predictions with stepwise adaptive task difficulty. The model recovers well- C Model predictions with linearly increasing task difficulty at a faster pace. Even-
established patterns of skill acquisition, performance, and engagement. Note that the tually the learner cannot keep up with the fast pace (C1), no longer recovers moti-
qualitative patterns shown here remain robust to variations in the main parameters vation (C3), and thus stops working entirely (C5). In all panels, parameter w = 0.5.
(see SI for details). B Model predictions with linearly increasing task difficulty at a

Within each period of constant task difficulty, behavior was similar to When the rate of increase in task difficulty is not overwhelming, there is
that found in the constant scenario, with plateaus in skill and performance as a gradual increase in peak motivation, peak fatigue, and the duration of
skill approached task difficulty. With repeated, discrete increases in task working periods (Fig. 4B3–B5). This indicates increased levels of engage-
difficulty, however, both skill and performance surpassed these plateaus ment. This contrasts with what was found in the constant scenario, where
(Fig. 4A1, A2). peak motivation, peak fatigue, and work period duration all decreased as
Our model predicts that when the task difficulty is raised adaptively, skill increased (Fig. 3C–E), and is distinct from the periodic changes in the
one’s performance exhibits plateaus, dips, and leaps, as shown in Fig. 4A2. stepwise scenario (Fig. 4A3–A5).
Whenever task difficulty increases after reaching a plateau, performance Both the stepwise and continuous scenarios can lead to sustained
first dips, then increases to exceed the previous plateau. This qualitative increase in skill learning and performance, but how far can one push the pace
pattern is in line with empirical studies of performance plateaus22,23. of task difficulty increase? We evaluated a special case in the continuous
Similar to the model’s predictions in the constant scenario, each epi- scenario where we increase the task difficulty at a very high rate (Fig. 4C). In
sode of task difficulty level starts with a spike in peak motivation, a spike in this scenario, individuals give up after failing to keep up with the ever more
peak fatigue, and longer duration of working periods. This suggests intense challenging task (Fig. 4C). Notably, on short timescales, this special case is
engagement as one first attempts to master a higher task level. Peak moti- characterized by a substantial increase in both peak fatigue (Fig. 4C4) and the
vation, peak fatigue, and working period duration all decline as skill duration of resting periods (Fig. 4C5). These short-timescale dynamics suggest
improves and approaches the target task difficulty, essentially recapitulating the task became too demanding, with the individual taking longer recover,
what we find in the constant scenario (Fig. 4A3–A5). Note that while we without sufficient motivation to compensate for the escalating exertion.
assume specific values for the timescale parameters c1 through c4 in our
simulations, the qualitative findings remain robust to variations in these Individual differences: learning-driven individuals overtake
parameters (see Supplementary Information for details). performance-driven individuals over the long run
We next analyze the impact of individual differences in motivation on the
Skill learning under continuous increases in task difficulty long-term development of skill and performance. Research in education and
To incorporate activities that are not characterized by stepwise change of social psychology has found that individuals differ in their relationship to
difficulty, we also examine a scenario where the task becomes increasingly challenging learning opportunities: While some individuals are intrinsically
more challenging in a continuous manner (e.g., a gradually increasing speed motivated to seek out new challenges, others prefer situations where they
of running and cycling (Fig. 4B, C). can demonstrate their existing mastery59–61. One study, for instance, found
We observe a qualitatively similar pattern in skill learning and per- that students who were more intrinsically motivated were more likely to
formance compared to that in the stepwise scenario—both scenarios induce choose a more challenging task, compared to students who were more
sustained increases in skill and performance over the long timescale, with competent but less intrinsically motivated60. We thus sought to simulate
the continuous scenario producing smoother changes. individual differences in learners’ source of motivation and choice of task.

npj Complexity | (2025)2:15 6

https://doi.org/10.1038/s44260-025-00039-x Article

Fig. 5 | Over the long run, learning-driven indi-

viduals achieve higher skill and performance than
performance-driven individuals. A Skill learning
in individuals motivated by skill (w = 1; solid line) or
performance (w = 0; dashed line). B Performance in
individuals motivated by skill (w = 1; solid line) or
performance (w = 0; dashed line). When individuals
are learning-driven, they are quicker to attempt
more difficult tasks, compared to performance-
driven individuals. They thus achieve a higher skill
and better performance over the long run.

We thus manipulated whether individuals were primarily motivated by of learning. By situating these dynamics within different training regimes—
learning (i.e., rate of increase of skill) or performance. According to Eq. (5), that is, different plans for increasing the task difficulty over time—we
learning-driven individuals (i.e., high w value) derive proportionally more investigated how people can achieve life-long learning or fail to maintain
motivation from their skill development (dS(t)/dt), while performance- their personal growth. Our approach thus offers a general, task-agnostic
driven individuals (i.e., low w value) derive proportionally more motivation framework for understanding the multiscale dynamics of learning across
from their rate of execution (E(t)). These two scenarios, therefore, capture varied domains.
two behavioral profiles that have been established in the literature on Our model does not distinguish between the kind of task (such as
intrinsic motivation and challenge-seeking. between mental and physical skills) and captures the broad-stroke phe-
To capture the finding that intrinsically-motivated individuals are also nomena in motivation, fatigue, and skill learning that transcend specific
more likely to seek out more challenging tasks60, we allowed ϵ—the activities. The model thus offers a formal account of why similar patterns
threshold for increasing task difficulty in our stepwise scenario—to depend recur across domains. This framework, however, can be modified to account
on whether individuals are motivated more by learning or performance: for other aspects of the learning process, including task-specific factors that
ϵ ¼ ϵmin þ kw, where ϵmin ¼ 0:005, and the slope k = 0.01. For simplicity, may play a role only within particular contexts (e.g., extrinsic encourage-
in exploring these individual differences, we chose a linear relation between ment from a coach or comrade), or distinct modes or types of learning that
source of motivation (w) and the threshold to seek out a more challenging may occur during different stages of the learning process (e.g., declarative vs.
task (ϵ), although more complicated nonlinear relationships are possible. procedural learning65). By formalizing key aspects of lifelong learning, we
As shown in Fig. 5, since performance-driven individuals are slower to offer a theoretical foundation on which other more elaborated accounts
attempt more difficult tasks, they suffer fewer dips in performance than can build.
learning-driven individuals. This reflects the learning-driven individual’s In our model, individuals are motivated in part by the gap between
drive to seek out more challenging tasks, even at the cost of short-term dips their current skill and the difficulty of the task (T − S). As a result, indivi-
in performance. The performance of learning-driven individuals, therefore, duals do not work at all if the task is too difficult (Fig. 4C1), they experience
may repeatedly drop below the performance of performance-driven indi- peak motivation when the task is appropriately difficult relative to their
viduals. This short term deficit, however, is overcome by the learning-driven current skill, and motivation decreases as task becomes relatively boring.
individuals’ improved skill, so that eventually learning-driven individuals This period of maximal motivation, moreover, is accompanied by longer
perform better than performance-driven individuals, even when the periods of sustained effort—despite high levels of accumulated fatigue. In
learning-driven individual is experiencing a dip in performance while the other words, when the task’s challenge is aligned with the individual’s skill—
performance-driven individual is at their peak. not too difficult (Fig. 4C1) but not too easy (Fig. 3)—then individuals
experience high motivation, the capacity to work for extended periods of
Discussion time, and the capacity to sustain their efforts in the face of high fatigue
Life-long learning involves non-linear change on multiple timescales. Past (Fig. 4, panels B3–B5).
research has established a number of classic patterns in these dynamics, These emergent attributes of our simulated work are hallmarks of deep,
patterns that strikingly recur across varied domains9,22,23,62. Motivation sustained engagement. They resemble features of the phenomenon of
waxes and wanes from moment to moment27, skill can develop gradually “flow,” a state of mind that typically occurs when a task is sufficiently
over days or years1,2, and performance can rise and dip as new challenges are difficult—not too difficult to make progress but also not too easy to offer a
attempted23. Our formal model offers a framework for explaining these challenge62. Flow experiences consist of a cluster of features, including
known patterns, including how dynamics on one timescale can shape or heightened motivation and an ability to work for long periods despite
generate dynamics on another. fatigue62,64. This pattern shares striking similarities with the emergent
Our model offers a parsimonious account of several important phe- dynamics of work, rest, and motivation in our model, particularly when the
nomena in learning. Learning and performance, for instance, are known to task difficulty was neither too challenging nor too easy (e.g., Fig. 4A, B). This
exhibit gradual increases, static plateaus, or short-term dips22,23,63. In our raises the possibility—admittedly speculative—that the minimal mechan-
model, these long-timescale behaviors emerge from short-timescale fluc- isms that are formalized in our model may partially explain why flow states
tuations in motivation, fatigue, and effort. Seemingly static plateaus in long- consist of the particular cluster of features with which they are associated62.
timescale learning—a recurring pattern in studies of skill While episodes of deep engagement, such as flow states, have tradi-
acquisition11,13,14,23—may thus reflect ongoing dynamic interactions between tionally been discussed as a phenomenon that emerges and disappears on
micro-level processes at short-timescales. Moreover, the peaks and troughs short timescales—emerging during a single session of work, or reflecting the
of motivation and fatigue—and thus the amount of time individuals are able fit between current activities and goals66—our model offers a framework for
to sustain their effort—reflect individual differences in whether one is situating the short-timescale dynamics of deep, sustained engagement
motivated by intrinsic growth (i.e., skill development) or extrinsic within the long-timescale dynamics of skill development, performance, and
achievement (i.e., performance)62,64. The model thus offers a formal account task selection. When task difficulty increases in discrete jumps, for instance,
of how individual differences in motivation may relate to long-term patterns individuals experience brief periods of deep engagement, which disappear as

npj Complexity | (2025)2:15 7

https://doi.org/10.1038/s44260-025-00039-x Article

their skill catches up to the demands of the task, at which point the task is no Our work offers two contributions. First, our model offers a formal
longer sufficiently challenging (Fig. 4A). When the difficulty of the task is mathematical framework to study the interaction of these timescales, and to
increased continuously, however, deep engagement can be sustained and understand the multiscale dynamics of learning as an integrated dynamical
even increased over time (Fig. 4B). While our minimal model does not speak system. The stability of long-term dynamics (e.g., skill plateaus), for
to the subjective, phenomenological features of deep, sustained engagement, instance, emerges from dynamic and intricate interactions on shorter
it offers a parsimonious account of the coupled dynamics of motivation, timescales, and the functional form of long-timescale learning curve
work, and fatigue on both short- and long-timescales. depends on the instantaneous rate of learning. The influence goes in the
The emergence of deep, sustained engagement in our model depends other direction, too—different long-timescale regimes for increasing task
on the source of an individual’s motivation. When individuals are intrin- difficulty, for instance, produce different short-timescale dynamics of
sically motivated by learning (i.e., w close to 1), they sacrifice short-term motivation and fatigue. While common sense suggests that there must be
gains in performance for the opportunity to improve their skill (Fig. 5). This interactions between processes on these different timescales, our model
leads to frequent dips in performance. Over time, however, the benefits of spells out some of those potential interactions.
long-term cumulative learning outweigh the costs of these temporary per- Second, our work brings into conversation different research areas that
formance dips induced at the onset of new skill acquisition. This is a phe- are typically investigated in isolation. Research on life-long mastery is sel-
nomenon that has been described empirically in skill acquisition ranging dom connected to the small improvements possible during a session of
from the discovery of the Fosbury Flop technique in the sport of high jump practice. The mastery achieved by virtuosos such as the artist Picasso or the
(initial adoption of the Fosbury Flop can lead to a drop in performance athlete Muhammad Ali is treated separately from the modest achievements
before the technique is mastered), to the development of prodigious of regular people. Nevertheless, learning occurs on all timescales, and is
memory for a span of random digits23. Since performance dips are followed common to all humans. This paper is an attempt to bridge these timescales.
by periods of increased challenge and deep engagement, our model captures The model has a number of limitations. First, since our goal was to
the notion of the so-called autotelic personality type, characterized by a capture domain-general patterns in skill acquisition, we have deliberately
disposition to actively seek challenges and deep engagement67. The long- ignored task-specific constraints on skill acquisition. While learning chess
timescale dynamics of learning thus depend on the moment-to-moment and learning to juggle may share high-level, qualitative patterns, they differ
dynamics of motivation and task selection. of course in countless ways. Future work should incorporate task-specific
Maximizing long-timescale learning, however, requires a careful bal- considerations—such as limitations of the body, or task-specific strategies—
ance. For individuals to experience sustained, long-term learning, they must to develop formal accounts of learning within specific domains. Second, our
engage in challenges above their current skill level, but only to a degree. If model does not currently account for interactions between different skills
individuals choose to maximize their current performance rather than seek and activities. Many contemporary workplaces, for instance, require indi-
out new challenges (such as the performance-driven individuals in Fig. 5), viduals to adapt to job requirements that change over time, such that
then their long-term learning will progress more slowly. But if task difficulty workers must replace old skills with new ones, or even develop multiple new
is increased too quickly, then learning cannot keep up and eventually an skills in parallel. Future work should explore the ways in which acquiring
individual will cease to learn, perform, and work (Fig. 4C). When the one skill may help, or even hinder, the acquisition of another, and how the
increase in learning is carefully calibrated to an individual’s ability (Fig. 4A, pursuit of multiple skills in parallel can be synergistic or detrimental.
B), then learning and peak motivation can both grow over long periods68–70. As stated in the introduction, our mathematical and computational
The decision to attack entirely new tasks, or the discovery of a more model answers the call for theory building in psychological science4 by
difficult but potentially more effective approach to an existing task, can offering a unified framework for theorizing about the multiple timescales on
produce learning curves that consist of piece-wise power laws22. Here, our which human learning unfolds. This is not a statistical model of a particular
model reproduces this long-timescale dynamic when the task difficulty is dataset. Instead, we offer a principled, formal account of a complex phe-
increased in discrete jumps (Fig. 4A), with each period of decreasing returns nomenon. In so doing, we hope our work encourages scientists working in
followed by a period of renewed rapid growth. In our model, however, each separate disciplines to start a collective conversation about the complex,
period of decreasing returns actually follows an exponential rather than multiscale dynamics of skill learning.
power-law learning curve, in line with work showing that individual
learning curves follow an exponential form14,20,71. As we show analytically in Methods
the Supplementary Information, this exponential decay in the rate of All methodological details necessary to replicate the model are described
learning emerges from our assumption that the rate of learning depends comprehensively in the section, “The mathematical model”. Additional
only on relative task difficulty (i.e., the difference between task difficulty and details on mathematical derivations are included in Supplementary
skill, T − S). In order for learning to follow a power law, the rate of learning Information.
must depend not only on the amount of current challenge but also on the
passage of time itself (e.g., decaying as a function of the number of practice Code availability
trials). The difference between exponential and power laws of practice, The code used in the simulation, as well as data generated from the simu-
therefore, reduces to whether the classic phenomenon of diminishing lations, are available as supplementary files.
returns depends only on changes in skill (i.e., diminishing as skill approa-
ches saturation), or on both skill and time (e.g., also diminishing with each Received: 21 July 2024; Accepted: 10 March 2025;
attempt20). We speculate, therefore, that the functional form of an indivi-
dual’s learning curve may depend on the individual and the task. In cases
where the rate of learning decays with time (e.g., with an individual who just References
tries less over time), then learning should follow a power law; in cases where 1. Newell, K. M., Liu, Y.-T. & Mayer-Kress, G. Time scales in motor
the rate of learning depends only on the current challenge, then learning learning and development. Psychol. Rev. 108, 57 (2001).
should follow an exponential curve. 2. Huys, R., Daffertshofer, A. & Beek, P. J. Multiple time scales and
In the following paragraphs, we briefly summarize our work and multiform dynamics in learning to juggle. Mot. Control 8, 188–212 (2004).
highlight our contributions. Overall, skill mastery is the outcome of pro- 3. Smith, L. B. & Thelen, E. Development as a dynamic system. Trends
cesses that operate on widely varying timescales. Motivation and fatigue can Cogn. Sci. 7, 343–348 (2003).
change rapidly in response to context and activity, decisions to work or rest 4. Guest, O. & Martin, A. E. How computational modeling can force
are based on fluctuations in motivation and fatigue, while long-term theory building in psychological science. Perspect. Psychol. Sci. 16,
improvements in skill and performance reflect days or years of work. 789–802 (2021).

npj Complexity | (2025)2:15 8

https://doi.org/10.1038/s44260-025-00039-x Article

5. Borsboom, D., van der Maas, H. L., Dalege, J., Kievit, R. A. & Haig, B. 30. Herlambang, M. B., Taatgen, N. A. & Cnossen, F. The role of motivation
D. Theory construction methodology: a practical framework for as a factor in mental fatigue. Hum. factors 61, 1171–1185 (2019).
building theories in psychology. Perspect. Psychol. Sci. 16, 756–766 31. Hopstaken, J. F., Van Der Linden, D., Bakker, A. B. & Kompier, M. A. A
(2021). multifaceted investigation of the link between mental fatigue and task
6. Smaldino, P. E. How to build a strong theoretical foundation. Psychol. disengagement. Psychophysiology 52, 305–315 (2015).
Inq. 31, 297–301 (2020). 32. Hastings, A. Timescales, dynamics, and ecological understanding.
7. Marquet, P. A. et al. On theory in ecology. BioScience 64, 701–710 Ecology 91, 3471–3480 (2010).
(2014). 33. Hastings, A. Timescales and the management of ecological systems.
8. Kadanoff, L. P. Scaling and universality in statistical physics. Physica Proc. Natl Acad. Sci. USA 113, 14568–14573 (2016).
A 163, 1–14 (1990). 34. Flack, J. C. Multiple time-scales and the developmental dynamics of
9. Ericsson, K. A., Krampe, R. T. & Tesch-Römer, C. The role of social systems. Philos. Trans. R. Soc. B 367, 1802–1810 (2012).
deliberate practice in the acquisition of expert performance. Psychol. 35. Wagner, L. M. & Clifton, S. M. Modeling the public health impact of
Rev. 100, 363 (1993). e-cigarettes on adolescents and adults. Chaos 31, 113137 (2021).
10. Gladwell, M. Outliers: The Story of Success (Little, Brown, 2008). 36. Haken, H., Kelso, J. S. & Bunz, H. A theoretical model of phase
11. Simon, H. & Chase, W. Skill in chess. Am. Sci. 61, 175–188 (1973). transitions in human hand movements. Biol. Cybern. 51, 347–356
12. Ericsson, K. A. & Lehmann, A. C. Expert and exceptional (1985).
performanceevidence of maximal adaptation to task constraints. 37. Richardson, M. J., Marsh, K. L., Isenhower, R. W., Goodman, J. R. &
Annu. Rev. Psychol. 47, 273–305 (1996). Schmidt, R. C. Rocking together: dynamics of intentional and
13. Fitts, P. M. Human Performance (Belmont, California: Brooks/Cole unintentional interpersonal coordination. Hum. Mov. Sci. 26, 867–891
Pub. Co., 1967). (2007).
14. Estes, W. K. The problem of inference from curves based on group 38. Turvey, M. T. Coordination. Am. Psychol. 45, 938 (1990).
data. Psychol. Bull. 53, 134 (1956). 39. Abrams, D. M. & Panaggio, M. J. A model balancing cooperation and
15. Newell, A. & Rosenbloom, P. Mechanisms of skill acquisition. competition can explain our right-handed world and the dominance of
Cognitive Skills & their Acquisition (1981). left-handed athletes. J. R. Soc. Interface 9, 2718–2722 (2012).
16. Anderson, J. R. & Fincham, J. M. Acquisition of procedural skills from 40. Yang, V. C., Abrams, D. M., Kernell, G. & Motter, A. E. Why are us
examples. J. Exp. Psychol. 20, 1322 (1994). parties so polarized? A “satisficing” dynamical model. SIAM Rev. 62,
17. Stafford, T. & Dewar, M. Tracing the trajectory of skill learning with a 646–657 (2020).
very large sample of online game players. Psychol. Sci. 25, 511–518 41. Leonard, N. E., Lipsitz, K., Bizyaeva, A., Franci, A. & Lelkes, Y. The
(2014). nonlinear feedback dynamics of asymmetric political polarization.
18. Gallistel, C. R., Fairhurst, S. & Balsam, P. The learning curve: Proc. Natl Acad. Sci. USA 118, 1–11 (2021).
implications of a quantitative analysis. Proc. Natl Acad. Sci. USA 101, 42. McChesney, G. The psychology of efficiency. J. Appl. Psychol. 1, 176
13124–13131 (2004). (1917).
19. Haider, H. & Frensch, P. A. Why aggregated learning follows the power 43. Soderstrom, N. C. & Bjork, R. A. Learning versus performance: an
law of practice when individual learning does not: Comment on integrative review. Perspect. Psychol. Sci. 10, 176–199 (2015).
Rickard (1997, 1999), Delaney et al. (1998), and Palmeri (1999). J. Exp. 44. Akizuki, K. & Ohashi, Y. Measurement of functional task difficulty
Psychol. 28, 392–406 (2002). during motor learning: what level of difficulty corresponds to the
20. Heathcote, A., Brown, S. & Mewhort, D. J. The power law repealed: the optimal challenge point? Hum. Mov. Sci. 43, 107–117 (2015).
case for an exponential law of practice. Psychon. Bull. Rev. 7, 185–207 45. Guadagnoli, M. A. & Lee, T. D. Challenge point: a framework for
(2000). conceptualizing the effects of various practice conditions in motor
21. Myung, I. J., Kim, C. & Pitt, M. A. Toward an explanation of the power learning. J. Mot. Behav. 36, 212–224 (2004).
law artifact: insights from response surface analysis. Mem. Cognit. 28, 46. MacIntyre, P. & Serroul, A. 11. Motivation on a Per-Second Timescale:
832–840 (2000). Examining Approach- Avoidance Motivation During L2 Task
22. Donner, Y. & Hardy, J. L. Piecewise power laws in individual learning Performance. In Dörnyei Z, MacIntyre P. & Henry A. (eds) Motivational
curves. Psychon. Bull. Rev. 22, 1308–1319 (2015). Dynamics in Language Learning. Bristol, 109–138 Blue Ridge Summit:
23. Gray, W. D. & Lindstedt, J. K. Plateaus, dips, and leaps: where to look Multilingual Matters. https://doi.org/10.21832/9781783092574-013
for inventions and discoveries during skilled performance. Cogn. Sci. (2014).
41, 1838–1870 (2016). 47. Aragonés, D., Eekhoff, A., Horst, F. & Schöllhorn, W. I. Fatigue-related
24. van der Maas, H. L. & Molenaar, P. C. Stagewise cognitive changes in technique emerge at different timescales during repetitive
development: an application of catastrophe theory. Psychol. Rev. 99, training. J. Sports Sci. 36, 1296–1304 (2018).
395 (1992). 48. Burdack, J., Horst, F., Aragonés, D., Eekhoff, A. & Schöllhorn, W. I.
25. Boetje, J. & van Ginkel, S. The added benefit of an extra practice Fatigue-related and timescale-dependent changes in individual
session in virtual reality on the development of presentation skills: a movement patterns identified using support vector machine. Front.
randomized control trial. J. Comput. Assist. Learn. 37, 253–264 Psychol. 11, 551548 (2020).
(2021). 49. Brehm, J. W. & Self, E. A. The intensity of motivation. Annu. Rev.
26. Mirzaei, M., Ghadiri, F., Arsham, S. & Rajabi, H. The effect of seven Psychol. 40, 109–131 (1989).
sessions of training program with massed and distributed practice 50. Martin, K., Meeusen, R., Thompson, K. G., Keegan, R. & Rattray, B.
frequency on the acquisition of cycling skills in autistic and healthy Mental fatigue impairs endurance performance: a physiological
boy in tehran city. Sci. J. Rehabilit. Med. 10, 1228–1243 (2022). explanation. Sports Med. 48, 2041–2051 (2018).
27. Katzell, R. A. & Thompson, D. E. Work motivation: theory and practice. 51. Palmer, L. The relationship between stress, fatigue, and cognitive
Am. Psychol. 45, 144 (1990). functioning. Coll. Stud. J. 47, 312–325 (2013).
28. Konz, S. Work/rest: Part II—the scientific basis (knowledge base) for 52. Coa, K. & Patrick, H. Baseline motivation type as a predictor of
the guide. Int. J. Ind. Ergon. 22, 73–99 (1998). dropout in a healthy eating text messaging program. JMIR mHealth
29. van der Steen, S., Steenbeek, H. W., Den Hartigh, R. J. & van Geert, P. uHealth 4, e5992 (2016).
L. The link between microdevelopment and long-term learning 53. Ondersma, S. J., Winhusen, T., Erickson, S. J., Stine, S. M. & Wang, Y.
trajectories in science learning. Hum. Dev. 63, 4–32 (2019). Motivation enhancement therapy with pregnant substance-abusing

npj Complexity | (2025)2:15 9

https://doi.org/10.1038/s44260-025-00039-x Article

women: does baseline motivation moderate efficacy? Drug Alcohol 71. Anderson, R. B. & Tweney, R. D. Artifactual power curves in forgetting.
Depend. 101, 74–79 (2009). Mem. Cognit. 25, 724–730 (1997).
54. Atkinson, J. W. Motivational determinants of risk-taking behavior.
Psychol. Rev. 64, 359 (1957). Acknowledgements
55. Anderman, E. M. Achievement motivation theory: balancing precision We thank the Santa Fe Institute (SFI) and the James S. McDonnell
and utility. Contemp. Educ. Psychol. 61, 101864 (2020). Foundation (JSMF) for jointly organizing the conference that sparked many
56. Massar, S. A., Wester, A. E., Volkerts, E. R. & Kenemans, J. L. of the ideas explored in this paper. We thank Hilary Skolnik (SFI) and Susan
Manipulation specific effects of mental fatigue: evidence from novelty Fitzpatrick (JSMF) for organizing these conferences, as well as all our
processing and simulated driving. Psychophysiology 47, 1119–1126 colleagues and employees at the Santa Fe Institute for their support.
(2010). Additionally, we thank the Omidyar Fellowship for its support in making this
57. Nijp, H. H., Beckers, D. G., Kompier, M. A., van den Bossche, S. N. & research possible.
Geurts, S. A. Worktime control access, need and use in relation to
work-home interference, fatigue, and job motivation. Scand. J. Work Author contributions
Env. Health 41, 347–355 (2015). M.L. led the conceptualization of the study, V.C.Y. led the simulation of the
58. Kumalasari, L. D. & Elgeka, H. W. S. The mediating role of intrinsic mathematical model, and all authors contributed to the development of the
motivation in the relationship between organizational support and model and wrote the paper.
quantitative workload and work-related fatigue. J. Sains Psikol. 13,
28–41 (2024). Competing interests
59. Dweck, C. S. & Leggett, E. L. A social-cognitive approach to The authors declare no competing interests.
motivation and personality. Psychol. Rev. 95, 256 (1988).
60. Inoue, N. Why face a challenge?: the reason behind intrinsically Additional information
motivated students’ spontaneous choice of challenging tasks. Learn. Supplementary information The online version contains
Individ. Diff. 17, 251–259 (2007). supplementary material available at
61. Abuhamdeh, S. & Csikszentmihalyi, M. The importance of challenge https://doi.org/10.1038/s44260-025-00039-x.
for the enjoyment of intrinsically motivated, goal-directed activities.
Personal. Soc. Psychol. Bull. 38, 317–330 (2012). Correspondence and requests for materials should be addressed to
62. Csikszentmihalyi, M. Flow: The Psychology of Optimal Experience Mingzhen Lu, Tyler Marghetis or Vicky Chuqiao Yang.
(Harper & Row, New York, 1990).
63. Gray, W. D. Plateaus and asymptotes: spurious and real limits in Reprints and permissions information is available at
human performance. Curr. Dir. Psychol. Sci. 26, 59–67 (2017). http://www.nature.com/reprints
64. Csikszentmihalyi, M. Beyond Boredom and Anxiety (Jossey-Bass
Publishers, 1975) https://books.google.com/books?id= Publisher’s note Springer Nature remains neutral with regard to
afdGAAAAMAAJ. jurisdictional claims in published maps and institutional affiliations.
65. Kim, J. W., Ritter, F. E. & Koubek, R. J. An integrated theory for
improved skill acquisition and retention in the three stages of learning. Open Access This article is licensed under a Creative Commons
Theor. Issues Ergon. Sci. 14, 22–37 (2013). Attribution-NonCommercial-NoDerivatives 4.0 International License,
66. Melnikoff, D. E., Carlson, R. W. & Stillman, P. E. A computational which permits any non-commercial use, sharing, distribution and
theory of the subjective experience of flow. Nat. Commun. 13, 2252 reproduction in any medium or format, as long as you give appropriate
(2022). credit to the original author(s) and the source, provide a link to the Creative
67. Baumann, N. Autotelic Personality. In: Peifer, C., Engeser, S. (eds) Commons licence, and indicate if you modified the licensed material. You
Advances in Flow Research. 231–261 (Springer, Cham, 2021) https:// do not have permission under this licence to share adapted material
doi.org/10.1007/978-3-030-53468-4_9. derived from this article or parts of it. The images or other third party
68. Chow, J.-Y., Renshaw, I., Button, C. & Davids, K. Effective learning material in this article are included in the article’s Creative Commons
design for the individual: a nonlinear pedagogical approach in licence, unless indicated otherwise in a credit line to the material. If material
physical education. In (eds Ovens, A., Hopper, T. & Butler, J.) is not included in the article’s Creative Commons licence and your intended
Complexity Thinking in Physical Education (Routledge, 2012). use is not permitted by statutory regulation or exceeds the permitted use,
69. Son, L. K. & Sethi, R. Metacognitive control and optimal learning. you will need to obtain permission directly from the copyright holder. To
Cogn. Sci. 30, 759–774 (2006). view a copy of this licence, visit http://creativecommons.org/licenses/by-
70. Headrick, J., Renshaw, I., Davids, K., Pinder, R. A. & Araújo, D. The nc-nd/4.0/.
dynamics of expertise acquisition in sport: the role of affective learning
design. Psychol. Sport Exerc. 16, 83–90 (2015). © The Author(s) 2025

npj Complexity | (2025)2:15 10