Absorption-Gated Allocation for Left-Behind Learners

Recent theory says that when help compounds, meaning that students who are already ahead gain more from each hour of tutoring, a planner should spend its budget on whoever is furthest behind. This "lift the bottom" rule is called Rawlsian allocation, and the theory says it beats "reward the most responsive" (utilitarian) allocation over long horizons. We argue that this advice can quietly fail for one large group: rural children whose parents have migrated to cities for work, known as left-behind children. For many of them, low scores are only half the picture. The other half is a missing channel through which help can take hold: an adult at home who can supervise practice, and enough trust in adults to engage with a tutor. If that channel is closed, extra tutoring is largely wasted. A rule that always serves the lowest scorer can then pour a small rural budget into students who cannot yet absorb it, while the channel itself keeps eroding.

The proposal is to model each student with two states, academic mastery and "absorptive support," where support gates how much a tutoring hour is worth. A planner can spend on either state. We will work out when Rawlsian allocation on mastery alone breaks down. We will then build an estimator that recovers absorptive support from ordinary learning logs, so the planner does not depend on sensitive family surveys. Finally, we will test whether support-first allocation beats both Rawlsian and fairness-constrained bandit baselines at the per-student budgets typical of rural schools. The resource gap between urban and rural schools is not a side topic here. It sets the budget regime, and the budget regime is where the ranking of policies is predicted to flip.

Core Claim

Policy Design in Long-run Welfare Dynamics proves that under a Matthew effect, Rawlsian allocation (Min-U) almost surely achieves higher long-run welfare than utilitarian allocation (Max-U). The result requires a "survival condition," a budget large enough to keep everyone from decaying to zero. It also uses a single scalar welfare per agent and a return function shared by all agents. Our claim is that once return depends on a second state that decays faster for left-behind children, the effective survival condition becomes much stricter. Two consequences follow. First, Rawlsian targeting on mastery alone can enter a "sink" regime at budgets that the one-state theory would call safe. Second, a policy that first raises support above a threshold and only then invests in mastery restores the long-run advantage.

The minimal model is the following. Student $i$ has mastery $x_i(t)$ and absorptive support $s_i(t)$, receives tutoring $a_i(t)\in\{0,1\}$ and a support action $b_i(t)\in\{0,1\}$ (for example, a regular mentor or teacher contact), and evolves as

$$x_i(t+1) = x_i(t) - g(x_i(t)) + a_i(t)\, f(x_i(t))\, h\big(s_i(t)\big), \qquad s_i(t+1) = s_i(t) - \delta_i + b_i(t)\, u,$$

subject to $\sum_i a_i(t) + c\, b_i(t) \le B$. The symbols mean the following. $f$ and $g$ are the return and decay functions from the welfare-dynamics paper. $h$ is a smooth threshold that stays near zero below a level $\tau$. $\delta_i$ is the per-step erosion of support, which is larger when the caregiver is an elderly grandparent who cannot manage home-school communication. $u$ is the gain from a support action, $c$ is its cost relative to a tutoring slot, and $B$ is the budget.

The one-state model is the special case $h \equiv 1$. The quantity we will characterize is the smallest per-student budget $B/N$ at which Min-U on $x$ still avoids collapse. We will describe how that threshold grows with the correlation between low $x$ and low $s$.

Why this intersection, and what each line of work contributes

Three kinds of prior work play different roles.

The first kind motivates the gap. Two descriptive studies document that rural interventions are absorbed poorly. Yang's study of the Double Reduction policy examines China's 2021 policy curbing homework and private tutoring in rural schools. It finds that grandparent caregivers cannot support digital home-school work (the "guardian gap"), and that fewer than half of more than 1,000 surveyed teachers supported the added after-school duties. Zhang's comparative analysis describes an administrative "performance trap," in which funds flow to urban schools that produce visible admission metrics. Neither study formalizes how absorption shapes the return on an intervention. That is the step we take.

Psychological work on left-behind children suggests the gate is threshold-shaped rather than linear. The Liaoning Cyberball study plays a game in which a child is excluded from a virtual ball toss, then measures retaliatory aggression. It finds left-behind children start with much lower trust (60.36 vs. 86.81 on the children's trust scale), and that trust reduces their aggression less than it does for other children. This is a constraint on how far we can lean on that evidence. The study measures aggression in 82 students from one province, not learning, so we treat it as a reason to test for threshold shape, not as evidence that the threshold exists in learning data.

The second kind supplies methods. For the dynamics and baselines we use the Min-U/Max-U analysis from Policy Design in Long-run Welfare Dynamics. We also use SoftFair, which shows that index-based restless bandits can leave nearly half of a population with no intervention. A restless bandit here means a scheduler that chooses which individuals to treat each round while every individual's state keeps changing whether treated or not. SoftFair avoids this starvation by sampling arms through a softmax, a probabilistic choice weighted by estimated value. Both methods assume one observed state with stationary dynamics.

For estimating the latent state we draw on knowledge tracing, which means models that predict a student's next answer from their past answers. DEKT shows that moment-to-moment emotions such as frustration and boredom change learning gains. Its emotions are short-lived and screen-level, whereas our $s$ is a slow, persistent state. DFCD supplies cold-start diagnosis of mastery for new students and unseen exercises. Our proposal does not build a new tracing model. It reads a new quantity, responsiveness to practice at a given mastery level, out of existing ones.

Research Program

The program has three parts, each of which can succeed or fail on its own.

The first part is analytical. In the model above we will derive the survival budget for three policies: Min-U on mastery, Max-U on mastery, and a gated policy that assigns support actions to students with $s<\tau$ before tutoring them. We will show how the gap between these survival budgets scales with $\text{corr}(x,s)$ and with the spread of $\delta_i$. We expect that the gated policy's advantage is zero when $s$ is uncorrelated with low mastery or when $\delta_i$ is uniform. That is a testable boundary, not an assumption.

We will also place a reported fiscal gap on the budget axis. City schools receive up to 2.4 times the per-school budget of county schools (Yang Wang et al.). We will locate where that ratio falls relative to the derived thresholds under calibrated parameters. This also speaks to the concentration debate. Wang et al. argue that spreading resources evenly dilutes them. Our model says that concentration can be correct, but on support-building for a subset of students rather than on the top performers.

The second part is an estimator. From response logs we will estimate each student's responsiveness: the learning gain predicted by a trained tracing model per practice attempt, conditioned on current mastery. A student whose mastery is low and whose responsiveness stays flat across many attempts is a candidate "gated" student.

We will test two properties of this estimate. The first is whether responsiveness has a threshold shape, meaning a bimodal or kinked relationship to persistent affect. The second is whether it is separable from mastery, or merely a noisy re-description of low mastery. On ASSISTments Challenge we will use the affect labels that DEKT relies on as an observable stand-in for $s$. Persistent negative affect over weeks, not single problems, is the proxy.

The third part is a small field validation. In partner rural schools we would administer the 17-item left-behind screening questionnaire and the children's trust scale alongside platform logs. We would then check whether estimated responsiveness tracks left-behind status and trust better than mastery does. This is correlational and modest in size. Its purpose is to establish whether the log-based proxy means anything for the population the theory concerns.

Baselines

The naive reference is equal allocation: every student receives the same tutoring share each period. Studies in the file argue that equal inputs mask functional inequality, but it remains the policy most schools actually run, so every other policy is reported relative to it.

Max-U on mastery gives each slot to the student with the highest predicted immediate gain, the "bang for the buck" rule of Policy Design in Long-run Welfare Dynamics. That paper finds it wins over short horizons. We therefore report results at both short and long horizons, so that Max-U is not unfairly judged only on the long one.

Min-U on mastery is the Rawlsian rule and the most direct competitor, because the file's strongest theory says it is optimal under a Matthew effect. No reported result was found on Lacuna for Min-U applied to educational or knowledge-tracing data. Its only reported empirical validation is on roughly 40,000 SIPP households, the U.S. Census Bureau's longitudinal income survey, used there as welfare proxies.

The strongest existing algorithmic method is SoftFair on a mastery-only state. On the CPAP adherence data (a clinical dataset on use of a sleep-apnea breathing device), it retains 93% of intervention benefit against 98% for a myopic policy, while the Whittle index policy starves nearly 50% of the population (SoftFair). We will also run a Whittle index policy on the augmented $(x,s)$ state with $s$ known. This is an oracle upper bound, and it separates "the right state matters" from "our policy is clever."

All policies share the same simulator, the same budget $B$, the same horizon, and the same number of random seeds. Any tunable parameter, such as SoftFair's fairness knob $c$ or our threshold estimate $\tau$, gets an identical grid-search budget on held-out simulated cohorts.

The ablation that isolates our key ingredient replaces the estimated responsiveness with the true $s$ in simulation, and with a plain left-behind flag. This tells us how much of any gain comes from measuring absorption rather than from simply prioritizing a demographic group.

Beating the baselines means three things together. At rural-regime budgets the gated policy must exceed Min-U and SoftFair on long-run mean and bottom-decile mastery. It must keep the never-served fraction at or below SoftFair's. And it must come within a stated margin of the oracle.

For the estimator, the tracing models serving as simulators must be competitive. The reported numbers are DEKT at AUC 0.8413 on ASSISTments Challenge against 0.8059 for DIMKT (DEKT), and IEKT as the top of 19 models on XES3G5M (XES3G5M). The file gives no absolute AUC for IEKT. We will reproduce these models rather than invent numbers.

Why not just…?

Why not just run Min-U, since theory already says it wins? This is the real threat. If low support and low mastery nearly coincide, then serving the lowest scorer and serving the least-supported student are the same action, and Min-U suffices. The theorem in Policy Design in Long-run Welfare Dynamics is silent on this because its agents have one state. Our analytical sweep over $\text{corr}(x,s)$, together with the separability test on ASSISTments affect data, settles the question directly. If responsiveness turns out to be a monotone re-description of mastery, we report that Min-U is adequate and the gating idea is unnecessary.

Why not just add a left-behind indicator and give those students higher priority? It is cheap and requires no model. However, the Liaoning study reports substantial variation in trust within groups, and trust moderates behavior differently for left-behind and other children (Cyberball study). That suggests the flag is a coarse proxy for $s$. The flag ablation measures this directly. If flag-priority matches estimated responsiveness, we should recommend the flag, which is also easier to explain and audit.

Why not just give a standard restless bandit or reinforcement learner the richer state? With $s$ observed, a standard Whittle policy on $(x,s)$ should recover most of what we propose, and that is why we include it as the oracle. The difficulty is that $s$ is not observed. Collecting it by survey adds work for teachers who, per Yang, are already overloaded, the same friction that undermines teacher-rating systems like DFRP. So the contribution under this alternative becomes the characterization of when the state matters, plus a survey-free estimator of it. We do not claim a better scheduler.

Why not just use an emotion-aware tracing model like DEKT? DEKT predicts answers better by using transient affect, but it predicts rather than allocates, and its states are measured at the level of single problems. We use it inside the simulator and as a source of the affect proxy. The question it cannot answer is where a limited budget should go.

Proposed Experiments

The first decisive check is in simulation. We will sweep per-student budget, $\text{corr}(x,s)$, and the spread of $\delta_i$, and record where Min-U on mastery loses to the gated policy. The prediction is a region at low budget and high correlation, together with no advantage when $\delta_i$ is uniform. If the gated policy wins everywhere, including where theory says it should not, we treat that as a bug.

The second check is the estimator on ASSISTments Challenge and XES3G5M. On ASSISTments, week-level persistent negative affect should predict flat responsiveness at matched mastery, with a kink rather than a linear slope. On XES3G5M, the gated subgroup should be a population distinct from the low-mastery population, as measured by overlap and by mixture-model fit. EdNet-KT1 serves as a transfer check using the T-DEKT affect representations.

The third check is semi-synthetic allocation. Trained tracing models serve as simulators for tutoring, and the support action is simulated with calibrated $u$ and $\delta$. All baselines above are compared on long-run mean mastery, bottom-decile mastery, never-served fraction, and short-horizon gain.

The fourth check is the field pilot. It tests whether estimated responsiveness correlates with left-behind status and trust-scale scores beyond what mastery explains. A null result here would mean that the log proxy does not measure the construct in this population, and we would report that plainly.