Splitting a Fixed Time Budget Between Priming and Activation

Many processes work in two steps: a gentle first step loosens a resistant structure, and a second, more targeted step does the real work. Biomass chemists do this when they soak corn stover in mild alkali so that ozone can later reach and break the lignin inside it. Teachers do it when students watch a lecture video at home so that the class hour can go to solving problems. In both cases the total time is fixed: reactor residence time in one, a semester or a contact hour in the other. In both cases the first step has no value on its own. Its value is how much it improves the second step, and doing too much of it backfires. Over-priming biomass breaks down the sugars you wanted to keep. Over-loading pre-class preparation leaves students who never arrive ready; more than 20% of students reported this in a flipped-classroom trial (Wu et al.).

This note proposes treating "how should a fixed time budget be split between priming and activation?" as one optimization problem with one model shared across both settings. The model describes a two-stage process in which priming raises accessibility with diminishing returns, causes damage that grows with priming time, and multiplies the effect of activation. The claim is not that students behave like corn stover. The claim is that this shared structure, used as the prior inside a Bayesian optimizer, does two things a generic optimizer cannot. It finds good splits with fewer experiments. More importantly, it predicts how the best split moves when the total budget changes, for example when a course loses a third of its contact hours. That second ability is what institutions facing credit-hour cuts actually need.

Core Claim

Let $T$ be the total time budget, $p$ the time spent priming, and $a = T - p$ the time spent activating. The outcome is $y$: test accuracy for learners, or enzymatic glucan digestibility for biomass. We model the outcome as

$$ y(p, a) = \big(1 - e^{-p/\tau}\big)\, g(a) + c\, g(a) - \delta\, p^{\gamma}, $$

The terms are defined as follows. $\tau$ is the time scale over which priming opens the structure. $g(a) = 1 - e^{-a/\kappa}$ is the saturating return to activation with time scale $\kappa$. $c$ is the efficacy of activation on an unprimed structure. $\delta$ and $\gamma > 1$ set how fast the cost of over-priming grows: carbohydrate degradation in biomass, overload and skipped preparation in students. The multiplicative first term carries the idea taken from the biomass literature. Alkaline swelling matters because it lets ozone diffuse into expanded capillaries, which it cannot do in raw stover (Fang et al.; Zhang et al.).

Because the four or five parameters do not depend on $T$, a model fitted at two budgets makes a falsifiable prediction about the best priming fraction $p^*/T$ at a third budget. The proposal uses this equation as the mean function of a Gaussian process. A Gaussian process is a flexible regression model that also reports its uncertainty. The optimizer then picks the next split to test by the upper-confidence-bound rule, which tries points that are either predicted to be good or poorly understood. What is new is the shared structured prior and the cross-budget extrapolation test, not the use of a Gaussian process.

Relation to Prior Work

The method's closest ancestor is Optimizing Instructional Policies. Lindsey et al. used a Gaussian process with upper-confidence-bound search and a Beta-Binomial model of student responses. With it they found an optimal presentation duration of 1.15 s for face-name pairs and a blocking-and-fading policy that reached 68.6% accuracy, against 60.9–66.6% for alternative policies. We reuse their observation model and search strategy. We differ in three ways. Their policy space was not a fixed budget divided between two interacting stages. Their surrogate had no structural prior. They never asked how the optimum moves when the budget changes.

The flipped-classroom and blended-lab literature motivates the question but does not answer it. Wu et al. ran a randomized crossover trial with a fixed pre-class load. Gains were not significant (84.96 vs 82.12, p=0.199; 91.63 vs 89.20, p=0.257), and many students were overloaded. Two other papers move theory out of protected contact time without measuring how much should move: the blended cell-biology lab redesign (Exploration of the Online and Offline Blended Teaching Model) and the three-stage big-data course (Hybrid Teaching Reform).

Two results constrain the design. Pannitto et al. found that compressing activities to 30–60 minutes turned them into puzzles without understanding. This implies a minimum useful activation time, which the saturating $g(a)$ must capture rather than assume away. Marciniak et al. found that adding short targeted definitions helped more than adding longer text (answer adequacy 47% to 53% to 55%). This supports modeling priming with diminishing returns.

On the biomass side, three lines of work show that priming pays off only through what it enables downstream: alkaline-then-ozone treatment of corn stover, dilute phosphoric acid on poplar tuned by orthogonal experimental design (Chen et al.), and fungal pre-boring of bamboo before carbonization (Zhang et al.). These studies tune conditions with factorial-style designs. None treats total residence time as a fixed budget, and none asks how the optimal split shifts with that budget. This connects to the file's open problem about kinetic mismatch between slow priming and fast activation.

Research Program

The work has three parts.

The first is a reusable optimizer. It runs the structured-prior Gaussian process on the one-dimensional split $p/T$, with budget $T$ as a second input. It uses a Beta-Binomial likelihood for learner data, following Lindsey et al., and a Gaussian likelihood for digestibility.

The second is an online human-subject study that follows Lindsey et al.'s Mechanical Turk protocol and reuses their category-learning task. Priming is passive study of labeled examples. Activation is retrieval practice with feedback. Total time is fixed at three budgets, for instance 6, 10, and 15 minutes.

The third is a bench study on corn stover. Priming is mild NaOH swelling and activation is ozonolysis, run at three fixed total residence times. The readout is enzymatic glucan digestibility, with furfural and 5-hydroxymethylfurfural (HMF) measured to check the over-priming term.

In both testbeds the optimizer is fitted on two budgets and must predict the best split at the third. That prediction is then checked directly. If the shared functional form fits both domains, the parameters become interpretable. The ratio $\tau/\kappa$ says whether a process is priming-limited or activation-limited, which is a compact diagnostic for deciding where saved time should go.

Baselines

The naive reference in teaching is spending the whole budget on activation with no separate priming. Its counterpart in biomass is ozonolysis of raw stover. Both are run at every budget on the same task or feedstock. Any split policy must beat these, or the premise fails.

The practitioner baseline is the static rule people use today. In teaching this is a fixed flipped split resembling Wu et al.'s design, roughly half the time on preparation. In biomass it is the single best condition from an orthogonal design in the style of Chen et al., given the same total number of experiments as our optimizer.

The strongest competitor is unstructured Gaussian-process search with upper-confidence-bound selection over $(p/T, T)$, exactly as in Lindsey et al. It is the established budget-aware optimizer for instructional policies. For a reference figure, the paper reports 68.6% validation accuracy for its best two-dimensional policy against 60.9–66.6% for alternatives. That number comes from a different policy space, so it serves as a sanity anchor for our replicated task rather than a target. For corn stover alkaline-ozone digestibility, there is no reported result on Lacuna, so the biomass baseline numbers will come entirely from our own runs. Both optimizers get the same number of participants or runs, the same kernel family, the same hyperparameter fitting procedure, and the same acquisition rule. The only difference is the mean function.

The key ablation removes the multiplicative interaction: priming and activation contribute additively, with everything else the same. This tests whether the "priming only matters through activation" structure is doing the work.

Beating the strongest baseline means two things. First, a lower prediction error for the optimal split at the held-out budget, with the gap larger than the bootstrap interval across participants. Second, a validated policy at that budget whose accuracy or digestibility is at least as good as the unstructured optimizer's, reached with no more experiments.

Why not just…?

Why not just run plain Gaussian-process search, as Lindsey et al. did? This is the real threat. Inside the budgets it has sampled, an unstructured Gaussian process may find the optimum as fast as ours, because the problem is low-dimensional. Our claim rests on extrapolation to a budget the optimizer has not sampled, where an unstructured model falls back to its prior and the structured one does not. The held-out-budget test settles this directly. If plain search predicts the new optimum equally well, the structured prior has not earned its place, and we will report the method as a parameter-reading tool only.

Why not just run an orthogonal or factorial design, as the biomass papers do? Factorial designs work well for screening many factors. For a single split variable crossed with budget, though, they spread runs evenly instead of concentrating them near the optimum, and they provide no model for budgets outside the grid. The equal-run comparison against the Chen-style design measures whether that difference matters in practice. Wet-lab runs are expensive, so even a modest saving in runs would be useful.

Why not just flip the classroom with a sensible fixed split? The evidence for fixed flipping is weak. Wu et al. found non-significant gains alongside overload. The fixed-split baseline at every budget tests whether one rule-of-thumb fraction is close enough. If the best fraction barely moves across budgets in either domain, this alternative wins, and we will say so.

Why not just fit a knowledge-tracing model to existing logs? Knowledge-tracing models predict whether a student answers the next question correctly from their history, and simple ones are hard to beat (simpleKT; pyKT). However, they model correctness sequences, not how time is allocated. The richest benchmark, XES3G5M, records only question start times and not time on task. Existing logs therefore cannot identify the priming–activation trade-off. This is why the proposal collects new data with time controlled by design.

Decisive Checks

The central check is the held-out-budget prediction in each domain, compared against plain Gaussian-process search and the additive ablation. A second check is whether the over-priming term tracks independent damage measures: learner-reported preparation difficulty and skipped priming in teaching, furfural and HMF levels in biomass. If $\delta$ fits accuracy or digestibility but does not track these damage measures, the model is fitting curves rather than describing the mechanism. A third check is whether one functional form fits both domains without domain-specific terms. A clean failure here would itself be informative: it would show that the structural analogy between pretreatment and pedagogy breaks down in practice, which the literature asserts but has never tested.