【Thought】System Theory and Formal Computation Theory Perspectives on "The Effect of Different Support Strategies on Student Affect"
Author Note: This post is a personal speculative reflection prompted by a close reading of Le Tallec, Prihar, and Käser (2025). The authors have produced a careful and methodologically innovative contribution to learning analytics by repurposing large-scale randomised experimental data to evaluate affective outcomes of support strategies alongside the cognitive outcomes that prior work had primarily examined. The theoretical and computational framings proposed in this note are exploratory extensions offered in a spirit of intellectual curiosity. They make no claims about the authors’ intentions or the scope of their original study. The authors of the original paper bear no responsibility for the speculative reasoning that follows, and any errors in the mathematical formalisations or philosophical interpretations are entirely my own.
Reference Paper: Le Tallec, J., Prihar, E., & Käser, T. (2025). The effect of different support strategies on student affect. In LAK ‘25: Proceedings of the 15th International Learning Analytics and Knowledge Conference, March 03–07, 2025, Dublin, Ireland. ACM. https://doi.org/10.1145/3706468.3706469
1. Reading the Paper as a Relational Dynamical System
The paper by Le Tallec et al. investigates how the availability of support strategies (hints, worked examples, and scaffolding) causally affects students’ affective states (boredom, concentration, confusion, frustration) during interaction with an online mathematics platform. The empirical findings are clear and, in places, counterintuitive: all three strategies improve performance, yet they produce divergent affective profiles. Hints increase confusion and reduce concentration. Examples reduce confusion. Scaffolding increases frustration. When video delivery is compared to text, confusion drops while frustration rises for scaffolded video support. When a student actually uses support, boredom reliably decreases and concentration reliably increases across all three types, a pattern absent in the mere-availability comparisons.
The authors’ own interpretive conclusion is apt: different support strategies likely operate through different psychological mechanisms. One strategy’s benefit in one affective dimension may arrive alongside a cost in another.
The experimental design is cross-sectional in its affect measurement, observing states at discrete moments and leaving the continuous or sequential evolution of learner states under repeated support interactions unmodelled. Reading the paper through a systems-theoretic lens reveals that what has been measured is a fragment of a richer closed-loop relational system between a learner and an instructional environment. This note attempts to articulate that system more formally, and to examine what its structure implies for both the interpretation of the existing results and the design of future adaptive systems.
2. The Hidden Feedback Loop
The process described in the paper can be sketched as a sequential loop:
Platform offers support → Student requests or receives support → Affective state shifts → Behaviour changes in clickstream → Platform observes behaviour → Platform offers the next support
In the current experimental design, the loop is broken after the second step. The platform records what happened, with its support policy assigned in advance and held fixed throughout. The design is therefore an open-loop experiment in which support is a fixed treatment and affect is a measured outcome. The causal inference is valid, and the system as a whole carries richer dynamics than the experiment makes visible.
The closed-loop reformulation takes the following form. Let $x(t)$ denote the latent learner state at time $t$, encompassing affective dimensions (boredom, concentration, confusion, frustration) and cognitive dimensions (current knowledge, working memory load, prior experience). Let $u(t)$ denote the support stimulus delivered at time $t$ (hint, example, scaffold, answer-only, or none, further parameterised by medium). Let $y(t)$ denote the observable behaviour emitted in the clickstream (time on task, attempt count, help requests, correctness). Because clickstream features record both the student’s internal processing and their interactions with the specific support options provided, the observation function takes both the latent state and the delivered support as arguments:
$$x(t+1) = f(x(t), u(t))$$
$$y(t) = h(x(t), u(t))$$
$$\hat{x}(t) = \phi(y(0), \ldots, y(t))$$
$$u(t+1) = \pi(\hat{x}(t))$$
Here $f$ is the state transition function, $h$ is the observation function mapping the latent state and the delivered support to observable behaviour, $\hat{x}$ is the estimated state produced by the Bi-LSTM affect detection model $\phi$, and $\pi$ is the support policy. In the current paper, the $f$ component is partially characterised through its average treatment effects on the affective coordinates of $x$. The policy $\pi$ is fixed and experimentally assigned. The observation model $h$ is implicit in the clickstream features selected for model training. Closing the loop requires all four components to be defined and coupled consistently, which corresponds precisely to the architecture of an intelligent tutoring system with real-time affect adaptation.
3. The Structure of the State-Input Coupling
Once the system is written as $x(t+1) = f(x(t), u(t))$, the central modelling question becomes: what is the structure of the coupling between the current state $x$ and the support input $u$?
A useful starting approximation is additive linear coupling:
$$x(t+1) = A x(t) + B u(t)$$
where $A$ is the autonomous dynamics matrix and $B$ is the input gain matrix. This form assumes each support type pushes the learner’s affective state in a fixed direction regardless of the learner’s current location in affect space. It is psychologically implausible for anything beyond the coarsest description, since it cannot represent the observation that hints may benefit students at low confusion levels while adding cognitive load for those already highly confused. The additive model also faces a structural constraint: if $x(t)$ is constrained to live on the probability simplex (since the predicted affect probabilities sum to one), then $A$ and $B$ must be chosen so that $x(t+1)$ remains on the simplex at every step. The unconstrained linear model generally fails to satisfy this requirement. A practical implementation would require either a projection step after each update or a reparameterisation in simplex coordinates, such as the log-ratio transform, before applying linear dynamics.
A considerably more expressive structure is bilinear coupling:
$$x(t+1) = \left(A + \sum_k u_{t,k} , B_k\right) x(t)$$
where $u_{t,k}$ is the activation weight of support mode $k$ and each $B_k$ reshapes the system’s local dynamics under that specific support. In this formulation, support modifies the matrix governing state evolution. The psychological content of this form is that each support type changes the geometry of how the learner’s state evolves, encoding the claim that different mechanisms are activated by different support types. The same simplex-preservation requirement applies: for the state to remain a valid probability vector, the combined operator $(A + \sum_k u_{t,k} B_k)$ must map the simplex to itself at every step. This is achievable, for instance, by requiring the combined matrix to be a column-stochastic matrix at all times.
A related structure is nonlinear gating, where the effective influence of support is modulated by the learner’s current state:
$$x(t+1) = f!\left(x(t),\ g(x(t)) \odot u(t)\right)$$
Here $g(x(t))$ is a state-dependent gating function and $\odot$ denotes elementwise modulation. This encodes the idea that the learner’s current affective configuration determines how strongly a given support acts: a student in sustained concentration may respond weakly to an example, while a student in high confusion may be strongly redirected by the same input. The gating function $g$ is in principle estimable from individual-level affect trajectory data.
A third candidate is a mode-switched hybrid system:
$$x(t+1) = f_{\sigma(u(t))}!\left(x(t)\right)$$
where the support input $u$ selects one of several distinct dynamical regimes, each with its own simplex-preserving map $f_\sigma$. The paper’s empirical findings are consistent with this form: hints, examples, and scaffolding produce qualitatively distinct affective trajectories, suggesting they select genuinely different modes of learner-environment coupling. The mode-switched model is practically attractive because each $f_\sigma$ can be estimated independently from the data of the corresponding experimental condition, making the parameter identification problem tractable.
4. Phase Space Geometry: Continuous Manifold and the Observation Functor
The dynamical system framing requires a careful treatment of what kind of space the learner’s affective state inhabits. The underlying affect space is a continuous geometric manifold $\mathcal{M}$, a smooth space in which the learner’s state varies continuously over time, driven by the coupled dynamics of cognition, emotion, and environmental input. This continuous manifold is the primary object. The discrete affective labels used in the paper (boredom, confusion, concentration, frustration) are neither elementary objects that constitute this space, nor coarse-grained versions of its points. They are objects in an entirely different category, reached through a partitioning operation on $\mathcal{M}$ and the application of an observation functor.
To make this precise, consider the learner’s true latent state $x(t)$ evolving on $\mathcal{M}$. The autonomous vector field $f(x, u)$ on $\mathcal{M}$ possesses attractors, regions toward which trajectories converge, and the boundaries between their basins of attraction define a topological partition of $\mathcal{M}$. This partition $\mathcal{P} = {R_B, R_N, R_U, R_F}$ assigns each point of the manifold to one of four regions. The affective labels are then the names of these regions: they are the image of the partition map $\pi: \mathcal{M} \to \mathcal{L}$, where $\mathcal{L} = {B, N, U, F}$ is a discrete set. The labels inhabit $\mathcal{L}$; the learner’s state inhabits $\mathcal{M}$. These are distinct spaces and the map between them is neither invertible nor structure-preserving in any metric sense.
In category-theoretic terms, one may formalise this as follows. Let $\mathbf{M}$ be a category whose objects are points of $\mathcal{M}$ and whose morphisms are continuous paths (trajectories). Let $\mathbf{L}$ be the discrete category whose objects are the four affective labels and whose only morphisms are identity morphisms. The observation operator $\Phi: \mathbf{M} \to \mathbf{L}$ is a functor that maps each manifold point to its label under the partition and each trajectory segment to the (trivial) morphism between the labels of its endpoints. This functor is many-to-one: all points within a basin $R_s$ map to the same label $s \in \mathbf{L}$. The discrete labels are objects in the target category $\mathbf{L}$, and their discreteness is a property of the target category, not of the source manifold. Saying the labels “coarse-grain” $\mathcal{M}$ would be to conflate the source and target of $\Phi$; the correct statement is that they are the image of $\Phi$, living in a categorically distinct space.
This logical ordering has a concrete consequence for interpreting the paper’s affect detection model. The model outputs a probability vector $\hat{x}(t) \in \Delta^3 \subset \mathbb{R}^4$, constrained to the three-dimensional simplex because the four predicted probabilities sum to one. With the Fisher information metric as its natural Riemannian structure, this simplex is a curved statistical manifold. It is the model’s probabilistic representation of the partition: rather than committing the estimated state to a single label in $\mathbf{L}$, it expresses a distribution over labels, encoding uncertainty about which basin the student’s true state currently occupies on $\mathcal{M}$. A student with substantial probability mass on both confusion and frustration simultaneously is not ambiguously assigned between two discrete states; their true state on $\mathcal{M}$ lies in a region whose image under $\Phi$ is genuinely uncertain, likely near a basin boundary. The simplex representation is therefore a probabilistic lifting of the functor $\Phi$ to a measure-valued map, and its interior is a representation of epistemic uncertainty about position on the continuous manifold.
Support acts on the continuous manifold as a perturbation of the vector field. When a support input $u$ is delivered, the flow on $\mathcal{M}$ changes: attractors may shift, basin boundaries may move, and the learner’s trajectory is deflected. The observable outcome in the label space $\mathbf{L}$, whether the student ends up classified as concentrated or frustrated, is determined by which basin the trajectory occupies after the perturbation, that is, by the image under $\Phi$ of the new trajectory segment. The paper’s regression coefficients estimate the mean change in the probability distribution over $\mathbf{L}$, which approximates the mean change in basin membership probability induced by the treatment. The deeper dynamical question concerns how the vector field $f(x, u)$ on $\mathcal{M}$ itself is reshaped by each support type, a question the functor $\Phi$ cannot answer because it discards the within-basin geometry entirely.
A student whose state on $\mathcal{M}$ lies deep within the concentration basin, far from any basin boundary, will be relatively robust to perturbation: even a large displacement in the vector field may not suffice to push the trajectory across a boundary and change its label. A student whose state lies near the boundary between the confusion and frustration basins will be highly sensitive to the same perturbation, with the trajectory tipping into one basin or the other depending on the local geometry of the support’s action. The paper’s aggregate effects average over all positions on $\mathcal{M}$, and the mean displacement in $\Delta^3$ will therefore significantly underestimate local effects near basin boundaries and overestimate them far from them.
The post-receipt transition results fit naturally within this picture. The act of requesting and receiving support functions as a perturbation strong enough to displace the trajectory across a basin boundary into the concentration basin, across all support types. This is the shared primary effect visible in the consistent boredom reduction and concentration increase. The support-specific secondary effects (examples and scaffolding reducing confusion and frustration; hints increasing them) reflect the residual trajectory dynamics after that basin-switching event, governed by how each support type reshapes the local vector field within the concentration basin.
5. Generative Grammar and Automaton Formalisations
The transition structure between affective states under different support types admits a natural formalisation as a controlled stochastic grammar or a probabilistic finite automaton. Let the affective states be the symbol alphabet $\Sigma = {B, N, U, F}$, standing for Boredom, coNcentration, confUsion, and Frustration respectively, using distinct letters to avoid the ambiguity of assigning $C$ to two different states. Let the support types be the input alphabet $\mathcal{U} = {h, e, s, a}$ (hint, example, scaffold, answer-only). A controlled Markov grammar assigns transition probabilities:
$$P(s’ \mid s, u)$$
where $s$ and $s’$ are the current and next affective states and $u$ is the support delivered. The paper’s results provide empirical estimates of the signed direction of these transitions at the population level. A full probabilistic parameterisation of $P(s’ \mid s, u)$ could in principle be extracted from individual-level affective state sequences, which the Bi-LSTM model could generate if applied to each student’s clickstream across time.
This formalism connects naturally to the theory of Partially Observable Markov Decision Processes (POMDPs). In a POMDP framing, the platform is the decision agent, support types are its actions, and the reward signal encodes some combination of learning gain and affective wellbeing. The true learner state $x$ is hidden; the platform observes only the estimated affect probabilities $\hat{x}$ produced by the detection model, which stand as a noisy proxy for the true state. The platform’s optimal support policy $\pi^*$ is the solution to the POMDP planning problem defined by the transition function $P(s’ \mid s, u)$, the observation function $P(\hat{x} \mid x)$, and the reward function. The paper provides, as a by-product of its causal analysis, population-level empirical data toward the estimation of $P(s’ \mid s, u)$, which is a foundational component of such a POMDP formulation.
6. Stimulus-Dependent Sensitivity and the Treatment Effect as a Local Approximation
One of the paper’s most conceptually rich findings is implicit in the divergence between support types: the sensitivity of the learner’s affective state to support input depends on both the current state and the specific input type. This is the formal content of the claim that different supports operate through different psychological mechanisms.
A note on precision is warranted here. The regression coefficient $\beta_z$ in each of the 48 regressions is a population-average treatment effect (ATE): the mean difference in predicted affect probability between students assigned to the treatment condition and those assigned to the control, averaged across all students and all positions in affect space. This is a population-level, linear approximation of the effect of changing $u$ on $x$. When $u$ is encoded as a structured vector (as discussed in Section 9), this approximation corresponds to an entry of the sensitivity matrix $\frac{\partial f}{\partial u}$ evaluated at the population-average state. For a categorical input, the partial derivative is understood as a finite-difference sensitivity, valid under the vector encoding but requiring explicit justification before being applied to categorical encodings directly.
With this clarification, the paper’s results suggest that the sensitivity matrix $\frac{\partial f}{\partial u}\big|_{\bar{x}, u}$ varies substantially across support types. For students near the population-average state $\bar{x}$, the sensitivity of the confusion coordinate to example support is estimated to be strongly negative (examples reduce confusion), while the sensitivity of the confusion coordinate to hint support is positive (hints increase cognitive load without resolving uncertainty):
$$\left.\frac{\partial x_{\text{confusion}}}{\partial u_{\text{example}}}\right|{\bar{x}} \ll 0 \qquad \left.\frac{\partial x{\text{confusion}}}{\partial u_{\text{hint}}}\right|_{\bar{x}} > 0$$
Video scaffolding produces a sensitivity vector with a strongly negative confusion entry and a positive frustration entry simultaneously, indicating that the same stimulus drives the state in constructive and costly directions within the affect simplex at the same time. This is the mathematical content of the empirical observation that a support can improve performance while increasing frustration: the performance coordinates and the affective coordinates of the state vector respond to different projections of the same input, and these projections are governed jointly by the coupling structure $f$.
These are population-average sensitivities. The state-dependent coupling models of Section 3 predict that the Jacobian $\frac{\partial f}{\partial u}\big|_{x, u}$ will vary with $x$, so that individual students in different regions of the affect simplex will experience the same support with qualitatively different local sensitivities. Estimating these region-specific sensitivities requires data at the individual level with sufficient state diversity, which the existing dataset does not yet provide at the required granularity.
7. Theory of Mind in the Support Relation
Support carries a dimension that purely dynamical models tend to abstract away. Each support type encodes, implicitly, a model of the learner’s current cognitive and affective state held by the instructional system or its designers. A hint presupposes: “I believe you can reach the solution with partial guidance.” Scaffolding presupposes: “I believe you need structured, step-by-step support to stay within your learning capacity.” An answer presupposes: “The current learning goal is continuation.” Each of these presuppositions constitutes an implicit Theory of Mind (ToM) attribution made by the support system about the learner. In human tutoring, this attribution is dynamic, bidirectional, and continuously updated through dialogue. In automated tutoring systems, it is fixed into the support policy at design time and refined only through whatever state estimation the platform performs.
The quality of the ToM embedded in the policy, measured by how accurately it models the learner’s actual cognitive and affective configuration, determines the quality of the support. Computationally, the support policy $\pi(\hat{x}(t))$ is an implicit ToM model: it maps an estimated learner state to a support action, and its accuracy is bounded by the accuracy of the affect detection model $\phi$ that produces $\hat{x}$.
This has implications extending beyond immediate affect. Repeated exposure to a particular support philosophy shapes the learner’s epistemic self-model over time. A student who consistently receives scaffolding may internalise the belief that problems are sequences of manageable steps, developing an analytical and systematic approach to difficulty. A student who consistently receives answers may develop a passive orientation toward knowledge acquisition. A student who consistently receives hints may develop a higher tolerance for uncertainty and a stronger habit of autonomous inference. In this sense, support participates in what might be called subject formation: the gradual construction of a learner’s identity as an agent in relation to knowledge and difficulty. The affective outcomes the paper measures are the most immediately observable traces of this deeper, longer-term process.
8. Quantum Lifting as a Formal Extension
The frameworks described above are grounded in classical probability and discrete-time stochastic dynamics. A more speculative extension considers representing learner states within a quantum formalism, motivated by the structural properties of superposition and non-commutativity rather than by any physical quantum mechanics of cognition.
In this extension, the learner’s state is represented as a density operator $\rho$ on a Hilbert space $\mathcal{H}$ spanned by basis vectors $|B\rangle, |N\rangle, |U\rangle, |F\rangle$ corresponding to the four affective states. A mixed state $\rho = \sum_{i} p_i |\psi_i\rangle\langle\psi_i|$ generalises the classical probability vector and allows off-diagonal coherence terms representing superposed affective tendencies.
One structural issue requires explicit attention. If support operations are modelled as unitary transformations $\hat{U}_u$ such that $|\psi’\rangle = \hat{U}_u |\psi\rangle$, the dynamics are norm-preserving and reversible. Affective transitions in real learners are almost certainly irreversible: frustration accumulated during scaffolding does not simply undo itself when the scaffolding is removed. A more appropriate formalism for irreversible evolution is the Lindblad master equation (continuous time) or its discrete-time analogue, the Kraus map:
$$\rho(t+1) = \sum_k E_k(u) , \rho(t) , E_k(u)^\dagger$$
where ${E_k(u)}$ are Kraus operators indexed by support type $u$, satisfying $\sum_k E_k^\dagger E_k = I$ to preserve trace and positivity of $\rho$. This open-system formulation can model both coherent evolution and decoherence (the decay of off-diagonal coherences under repeated interaction), and it subsumes the unitary case as a special instance.
A second point requiring disambiguation concerns the nature of affective indeterminacy. The quantum formalism is most interesting when affective states before support are genuinely superposed tendencies, which is a stronger claim than saying the observer is uncertain about which classical state the student is in. Classical epistemic uncertainty corresponds to a diagonal density matrix; genuine quantum superposition corresponds to off-diagonal coherence terms that produce interference effects. Whether affective states exhibit such genuine coherence in any operationally testable sense is an open question. The Bi-LSTM model’s output is a probability distribution over states, which corresponds to a diagonal density matrix and carries no information about coherences. Introducing off-diagonal terms would require an argument that learner states are superposed in a way that produces interference between affective trajectories, which remains at the level of structural analogy.
One property that the quantum formalism genuinely captures beyond classical probability is the non-commutativity of support sequences. If $\mathcal{E}_h$ and $\mathcal{E}_s$ are the Kraus channels for hint and scaffold respectively, then in general:
$$\mathcal{E}_s \circ \mathcal{E}_h (\rho) \neq \mathcal{E}_h \circ \mathcal{E}_s (\rho)$$
meaning the order in which supports are delivered may influence the final affective state in ways a commutative classical model cannot represent. Whether this non-commutativity is empirically significant in educational settings is a testable question the current paper does not address.
9. Encoding Support as a Control Signal
A practical question underlying all of the above is how support should be encoded as the input $u$ in the dynamical model. The choice of representation determines both the model’s expressive power and the validity of the sensitivity analysis developed in Section 6.
A purely categorical encoding, $u \in {\text{hint}, \text{example}, \text{scaffold}, \text{answer}}$, treats each support as an atomic symbol with no internal structure. This is the encoding implicit in the paper’s current analysis and is appropriate for a first-pass causal estimate. It does not represent gradations within a type (a detailed hint versus a minimal one, a scaffolded problem with two sub-steps versus six) or structural similarities across types (both examples and scaffolding reduce confusion, suggesting a shared component in their affect-space action), and it does not support partial-derivative sensitivity analysis directly.
A structured vector encoding represents each support as $u = (\text{type}, \text{medium}, \text{depth}, \text{timing}, \text{content complexity})$, enabling the partial-derivative interpretation of Section 6 to be applied consistently and allowing fine-grained analysis of which support dimensions drive which affective outcomes.
A semantic embedding represents the textual content of a support message as a vector in a high-dimensional semantic space via a language model encoder, enabling generalisation across supports that share meaning while differing in surface form.
Most natural for dynamical modelling is a geometric encoding that represents support directly as a displacement vector in the affect simplex, specifically as the mean shift in affect probabilities that delivery of a given support induces in a given region of the simplex. Under this encoding, the paper’s regression coefficients $\beta_z$ for each support condition provide direct empirical estimates of the support’s geometric action at the population-average state. The policy $\pi$ can then be formulated as a geometric optimisation problem: given the current state $x(t)$ on the simplex, select the support $u$ whose estimated displacement most efficiently moves $x$ toward the target region (high concentration, low frustration), subject to the simplex-preservation constraints discussed in Section 3.
10. Synthesis: Toward a Formal Model of Supportive Learning Dynamics
Drawing together the considerations above, the paper by Le Tallec et al. provides empirical grounding for a formal model of the following structure. The learner exists as a state $x(t)$ on the probability simplex $\Delta^3 \subset \mathbb{R}^4$, endowed with the Fisher information metric as its natural geometry. The instructional platform delivers support stimuli $u(t)$ drawn from a structured input space. The learner’s state evolves according to a state-input coupled dynamics $x(t+1) = f(x(t), u(t))$ in which the coupling is likely bilinear or nonlinearly gated, and in which the dynamics must be constrained to preserve the simplex. The state is partially observed through clickstream behaviour $y(t) = h(x(t), u(t))$, and an affect detection model provides a probabilistic estimate $\hat{x}(t)$ of the latent state.
The paper estimates, at the population level, the average treatment effects of support on each affective coordinate, which serve as linear approximations of the local sensitivity of $f$ with respect to $u$ evaluated at the population-average state under a vector encoding of support. These estimates reveal that the sensitivity is support-dependent and that different supports induce qualitatively different displacements in the affect simplex. The simplex has a basin structure shaped by support type, and individual learners occupy different regions of this simplex, suggesting that optimal support selection depends on current state position as much as on support type.
The closed-loop system the paper implicitly points toward, in which the platform’s support policy is dynamically adapted based on real-time state estimation, can be formalised as a POMDP, a switched dynamical system, or a controlled Markov grammar, depending on the level of mathematical granularity required. The ToM embedded in support design shapes affect in the immediate session and may shape learner identity over longer timescales. The quantum formalism, treated carefully as an open-system dissipative model with Kraus operators, offers a structurally coherent extension for representing non-commutativity of support sequences and the superposition of affective tendencies, with the caveat that empirical evidence for genuine quantum coherence in affect transitions remains absent.
The paper, read through these lenses, offers a first empirical characterisation of a dynamical system whose full formal structure remains to be articulated. That characterisation is valuable precisely because the causal identification it achieves through randomised assignment in authentic educational settings provides the kind of clean empirical input that formal dynamical modelling requires and rarely obtains.
This note was developed through a close reading of Le Tallec et al. (2025) and draws on dynamical systems theory, control theory, formal language theory (stochastic grammars and POMDPs), information geometry on the probability simplex, open quantum system theory (Lindblad/Kraus formalism), and the philosophy of mind and education. All mathematical formalisations are speculative proposals for future theoretical development and carry no claim of empirical validation in the educational context described here.