Compressing Homework Without Losing Learning: Time-Budgeted Practice Coresets

Many secondary-school students in China give up sleep to finish schoolwork. National surveillance found that nearly half of 13–17-year-olds miss sleep guidelines on school days (Huang et al.). A three-year cohort found sleep deprivation exceeding 96% by ninth grade (Dai et al.). That decline continued even under the "Double Reduction" policy, which capped homework and after-school tutoring, because families refilled the freed time with private study. The lesson is that a time cap alone does not hold. Nobody gives up practice they believe is necessary.

A second literature offers a way to show that some of that practice is not necessary. Representative subset selection, or "coreset" construction, picks a small subset of a large collection that preserves what matters about the whole. Seed banks keep about a sixth of their accessions and still retain over 94% of the alleles (Zhong et al.). Adaptive tests reach the same ability estimate with 15% fewer questions (Zhuang et al.). We propose to build and test a method that picks, for each student, a homework subset that fits a fixed nightly time budget. The subset is chosen to maximize the student's predicted learning, not the precision of a test score. Each subset comes with an estimate of how much predicted mastery is lost relative to the full assignment, which gives families something concrete to trust. The headline question is empirical. Can typical assignments be cut by a large fraction with no measurable loss on later performance? And does a learning-aware selection rule beat simple rules such as "a couple of items per skill"?

Core Claim

Existing educational subset selection optimizes the wrong thing for homework. BECAT (A Bounded Ability Estimation for Computerized Adaptive Testing) treats a test as a coreset problem. It picks questions whose loss gradient with respect to the student's ability matches that of the full question bank, so the subset measures ability as well as the whole bank would. It also assumes ability stays fixed during the session.

Practice inverts both assumptions. Ability is supposed to change during homework. An item that adds no measurement information, such as a fifth near-duplicate fraction problem, may still add consolidation, or it may add nothing. Which of these holds is an empirical property of the student and the skill, not a geometric fact about the item bank.

Our claim is that a homework coreset should be chosen by predicted learning gain under a time budget. That prediction should come from a knowledge-tracing model, meaning a model that predicts a student's future answers from their answer history. We expect this rule to compress more aggressively than measurement-matching or redundancy-pruning rules at equal downstream performance. If it does not beat simple skill-stratified sampling, that negative result is itself the useful finding: schools could adopt the simple rule.

Research Program

The selection problem takes four inputs: an assignment $Q$, a student with practice history $h$, a time budget $T$ in minutes, and per-item expected durations $t_i$ estimated from log timestamps. We solve

$$\max_{S \subseteq Q,\ \sum_{i \in S} t_i \le T} \; \sum_{k} w_k(h)\, \big[\hat m_k(h, S) - \hat m_k(h, \varnothing)\big]$$

Here $\hat m_k(h, S)$ is the knowledge-tracing model's predicted probability that the student answers a held-out item on skill $k$ correctly a week later, after practicing subset $S$. The weight $w_k(h)$ upweights skills on which this student is currently weak.

The weights encode a tension the file documents across fields. Distance-minimizing selection on phylogenetic trees (Gallagher et al.) deliberately avoids isolated outliers, because there they are usually noise. Priority sampling in garlic germplasm (Zhong et al.) deliberately keeps extreme accessions, because they carry rare useful traits. In homework, a student's weak skill is the "rare allele" and should not be pruned as noise. Cui et al. supports this weighting: they found the benefit of physical activity on cognition was over three times larger for the bottom 10% of students (0.394 vs 0.118). That result is about exercise, not homework, so we use it only as motivation to test heterogeneity, not as evidence that the weighting works.

If predicted gains show diminishing returns in $S$, the objective is approximately submodular. In that case, greedy selection by gain per minute, the standard knapsack-style greedy, is a reasonable solver, and we will check diminishing returns empirically rather than assume it.

The method has three parts. The first is a gain model. We fit a knowledge-tracing model on response logs and validate that its counterfactual predictions track reality. We do this by exploiting natural variation in how many items students actually completed on an assignment before a later item on the same skill. The second is the selector itself, plus an uncertainty estimate: the predicted mastery gap between $S$ and full $Q$, with bootstrap intervals over the gain model, reported as the "certificate." The third is a secondary analysis of whether logs show reduced per-item learning late in long sessions or at late clock times. Such a pattern would mean the last items of a long assignment are the cheapest to cut. It would also connect the method to the sleep literature, which so far rests only on self-report (Sha et al.).

The sleep direction plays two roles here. It sets the constraint, since the budget $T$ is chosen to protect sleep and exercise time. It also sets the outcome of the field pilot, which measures sleep and study time. Prior evidence motivates protecting non-study time. Sha et al. found that spare-time exercise predicts later sleep quality and not the reverse. Liu et al. found an inverted-U relation between exercise duration and achievement, with cognitive ability carrying about 91% of the mediated effect. Neither study shows that cutting homework helps, which is exactly what the pilot must test.

Baselines

The naive reference point is the full assignment, alongside random subsampling and time-truncation (do items in order until the budget runs out) at the same minute budget. These tell us whether any selection is needed at all. In germplasm, random sampling was consistently worse than priority sampling at preserving trait variance (Zhong et al.), but whether that carries over to practice items is unknown.

The strongest simple competitor is skill-stratified sampling. It takes a fixed number of items per tagged skill, which is the homework analogue of stratified core-collection construction (Ren et al.). We also run a per-student mastery-stopping rule: drop remaining items on a skill once the knowledge-tracing model predicts mastery above a threshold.

The strongest existing coreset method is BECAT (Zhuang et al.), repurposed to select homework items by gradient matching against the full assignment. Its reported results are 15% fewer questions at matched estimation accuracy across ASSIST, NIPS-EDU and EXAM, and a 2.48% AUC gain over Fisher-information selection on NIPS-EDU. Those numbers concern test measurement. For homework compression measured by later learning, we found no reported result on Lacuna, so this proposal would supply the first one.

Fairness rules apply to all methods. Every method receives the same logs, the same knowledge-tracing model where it needs one, the same skill tags, and the same minute budget. Every tunable method (BECAT, the stopping threshold, our weights) gets an equal hyperparameter search on a validation split. Two ablations isolate our ingredients: the learning-gain objective with uniform weights, which removes the priority term, and BECAT's measurement objective with our time budget and weights, which swaps only the objective. Beating a baseline means one of two things. Either we match full-assignment held-out performance, within a pre-registered non-inferiority margin, at a smaller budget than the baseline needs. Or we achieve significantly higher held-out accuracy at equal budget, with the gain concentrated in weaker students.

Why not just…?

Why not just cap homework time? Dai et al. (Sleep patterns among middle school students) show that a statutory cap failed to stop the sleep decline because families refilled the time. A cap says nothing about which work matters. Our pilot measures total evening study time directly to test whether a certified compressed set reduces refilling. If it does not, the certificate idea fails, and we will report that.

Why not just sample a couple of items per skill? This is a real threat. If tags are fine-grained and items within a skill are near-interchangeable, stratification may capture most of the available savings. The offline comparison at matched budgets settles the question. We expect our rule to win mainly where students differ in prior mastery, because stratification gives every student the same allocation. If the gap is within noise, we will recommend stratification and keep the pilot unchanged except for the selector.

Why not just use mastery-based stopping? Stopping is greedy within each skill and ignores the shared time budget across skills. It also inherits the cold-start problem that BECAT's authors note for early ability estimates. The ablation that keeps our gain model but replaces the budgeted selection with a stopping threshold shows whether the cross-skill budgeting adds anything.

Why not just apply BECAT directly? BECAT preserves the precision of an ability estimate under a fixed-ability assumption. Homework is meant to change ability. The objective-swap ablation tests this directly. If gradient matching performs as well on later held-out performance, our central claim is wrong.

Proposed Experiments

The offline study uses the three benchmarks BECAT used: ASSIST, NIPS-EDU (over 27,000 Eedi questions), and EXAM. Each is split by time so that every method selects from earlier responses and is scored on later responses to the same skills. The primary metrics are AUC and accuracy on those later responses as a function of the minute budget, reported separately for the bottom and top prior-mastery quartiles. The decisive offline check is gain-model validity. On students who naturally completed fewer items, the model's predicted loss from skipping must match their observed later performance. If it does not, no selector built on it can be trusted, and we stop before deploying.

The field study is a cluster-randomized pilot at the classroom level. One arm receives full homework, one receives stratified-compressed homework, and one receives our compressed homework with its certificate shown to parents, all at a budget chosen to protect sleep. Outcomes are next-unit test scores (with non-inferiority as the primary hypothesis), Pittsburgh Sleep Quality Index scores, self-logged total study time, and wrist actigraphy on a subsample, which addresses the self-report limitation in Sha et al..

The single most informative result would be twofold. Test scores would be non-inferior in both compressed arms, while sleep improves only in the certified arm. That would show the certificate, not the compression alone, is what stops time from being refilled.