Bayesian Simulated Students

• Bayesian simulated students are computational models that formalize learning as sequential Bayesian inference over structured concept spaces.
• They explicitly incorporate prior beliefs, noise models, and response-generation mechanisms to simulate student misconceptions and response variability.
• Evaluation methods include simulated teacher interactions, AUC-based learning curve analysis, and human-in-the-loop assessments to validate teaching approaches.

Bayesian simulated students are computational models that formalize the process by which learners acquire knowledge or skills under uncertainty, characterizing learning as sequential Bayesian inference over a well-defined concept space. These models are employed both as proxies for human learners in the evaluation of automated teaching strategies and as tools to generate large-scale ground-truth datasets for pedagogical and assessment research. The defining feature of Bayesian simulated students is the explicit specification of prior beliefs, noise models, and response-generation mechanisms, allowing systematic interrogation of teaching methodologies and validation against human behavior (Ross et al., 2024).

1. Mathematical Foundation

Bayesian simulated students encode a learner’s knowledge and misconceptions as probability distributions over concept spaces. Denote $H$ as the set of possible concepts—such as functions, generative model parameters, or programmatic procedures. A latent “student‐type” parameter $\alpha$ modulates their prior, encoding systematic misconceptions or biases representative of distinct student subpopulations. The initial belief is specified by the joint prior $p(\alpha)\,p_S(h\mid\alpha)$, where $p(\alpha)$ is uniform over student types, and $p_S(h\mid\alpha)$ is the student’s concept prior given their type (Ross et al., 2024).

Upon presentation of an example $(x, y)$, the model updates its beliefs using Bayes’ rule. For a finite output space $Y$, the "noisy-oracle" likelihood is given by:

$$p_S(\hat y \mid x, h) = (1-\epsilon)\, \mathbf{1}\{\hat y = h(x)\} + \frac{\epsilon}{|Y|-1}\,\mathbf{1}\{\hat y \neq h(x)\}$$

Here, $\epsilon$ models response noise, reflecting imperfect learning or guessing. Updates proceed according to the formula:

$$p_S(h \mid D_n, \alpha) = \frac{p_S(h\mid\alpha)\prod_{i=1}^n p_S(\hat y_i \mid x_i, h)} {\sum_{h'\in H} p_S(h'\mid\alpha)\prod_{i=1}^n p_S(\hat y_i \mid x_i, h')}$$

where $\alpha$0 denotes the sequence of prior examples and guesses.

2. Domains and Student-Type Parameterization

Bayesian simulated students have been instantiated for various educational domains, each with detailed concept encodings and student-type variants:

• Fraction arithmetic: Concepts correspond to compositional programs implementing fraction addition/multiplication via combinations of subroutines (nine in total). Two student types are defined:
  • add-learner: Overweights a spurious "multiplication-by-numerators-only" subroutine.
  • mult-learner: Overweights a spurious "addition-on-both-terms" subroutine.
• English past-tense morphology: Concepts are Naïve Bayes classifiers over four morphological classes; student types differ in prior knowledge about class distributions, with “unknown” classes initialized to uniform Dirichlet(1) and others using modes fit to Unimorph data.
• Function learning: Concepts are simple programs parameterized by predicates $\alpha$1, slopes $\alpha$2, and offsets $\alpha$3 (e.g., $\alpha$4). Student types are distinguished by possessing correct or spurious instantiations of $\alpha$5 or $\alpha$6 (Ross et al., 2024).

Priors are constructed using primitive counts, with large multipliers (e.g., $\alpha$7 for fractions) to ensure initial low (but non-negligible) credence in the target concept.

3. Response Generation and Simulation Protocol

Given a teacher prompt $\alpha$8 and interaction history $\alpha$9, Bayesian simulated students generate responses $p(\alpha)\,p_S(h\mid\alpha)$0 by sampling from their posterior-predictive distribution:

$p(\alpha)\,p_S(h\mid\alpha)$1

This method captures both parameter uncertainty and inherent stochasticity via the noise model.

Stepwise simulation proceeds by updating the posterior distribution over $p(\alpha)\,p_S(h\mid\alpha)$2 for each example-response pair, repeatedly drawing predictive responses for subsequent teacher queries. All hyperparameters (noise $p(\alpha)\,p_S(h\mid\alpha)$3, prior weights $p(\alpha)\,p_S(h\mid\alpha)$4, Dirichlet parameters) are fixed ex ante based on pilot data to align with observed human error rates and prior knowledge levels (Ross et al., 2024).

4. Evaluation and Validation Methodologies

Bayesian simulated students are used for both automated and human-in-the-loop evaluation:

• Simulated-teacher evaluation: Candidate teaching algorithms (e.g., AToM, GPT-4, non-adaptive baselines) interact with simulated students over fixed trajectories (40 steps for fractions and functions; 50 for verbs). Metrics include the posterior probability mass assigned to the true concept $p(\alpha)\,p_S(h\mid\alpha)$5 at each step and the area under the learning curve (AUC).
• Critical example selection: Tracks the fraction of early teaching examples that efficiently disambiguate student misconceptions from the target concept.
• Human-student evaluation: In tasks like function learning with the “wug machine,” human participants interact with adaptive teachers. Performance is measured via an AUC-style correctness score and the accuracy with which the teacher infers the participant’s student type (Ross et al., 2024).

This protocol enables comparison of teaching algorithms under controlled learner behaviors and experimental replication.

5. Inference Procedure and Implementation

Maintaining the student’s belief state is operationalized through vectorized inference:

$p(\alpha)$0 For each step, $p(\alpha)\,p_S(h\mid\alpha)$6 holds the posterior over hypotheses $p(\alpha)\,p_S(h\mid\alpha)$7 conditioned on history and student type $p(\alpha)\,p_S(h\mid\alpha)$8. The predictive distribution for future responses is computed as a weighted sum over this posterior.

When employed as a teacher model (e.g., AToM), there is an additional maximization over candidate $p(\alpha)\,p_S(h\mid\alpha)$9 to infer the most likely student type, using observed response histories.

6. Relation to Broader Bayesian Student Modelling

Bayesian simulated students are closely related to broader probabilistic models of student performance, such as those used in peer grading scenarios. In those settings, full Bayesian generative models account for latent ground-truth grades, grader noise, bias, and effort/spam variables; inference is carried out via Gibbs sampling, and output can be rounded and explained via mixed integer programming (Zarkoob et al., 2022).

While these peer-grading models focus on the reverse problem (pooled estimation of truth and student reliability), Bayesian simulated students as formulated for adaptive teaching concentrate on forward-propagating learner uncertainties, capturing misconceptions and response trajectories under pedagogically relevant noise structures. Both frameworks leverage the capacity of Bayesian inference to simulate, predict, and retroactively assess learning and grading dynamics in educational ecosystems.

References:

• "Toward In-Context Teaching: Adapting Examples to Students' Misconceptions" (Ross et al., 2024)
• "Better Peer Grading through Bayesian Inference" (Zarkoob et al., 2022)