Temperature controls
Choose exact values or drag the sliders. Paper default: 0.10 for both parameters.
Wrong answer → misconception prototypes
Temperature controls how strongly a wrong answer is associated with the most similar shared prototypes.
Student state → wrong-answer probabilities
Candidate scores depend on the student misconception state and answer associations; τans controls the final probability sharpness.
The scores above are m⊤αq,a before the τans Softmax. Changing τmis also changes these scores.
Beyond temperatures
Adjust the prototype-space capacity or the joint-loss weighting. These controls illustrate structure and objective terms, not predictions of retrained-model accuracy.
Number of misconception prototypes
M sets the number of shared prototype vectors and dimensions in each student's misconception state.
Changing M changes the model architecture and requires retraining. The five-prototype temperature example above is kept fixed for clarity.
CWP loss weight
λ balances the correctness and CWP objectives during training; it does not directly rescale inference probabilities.
Fixed illustrative losses: Lkt = 0.50 and Lcwp = 0.80
The percentages show shares of this illustrative scalar loss, not shares of gradient magnitude or measured training performance.
What τmis changes
It sharpens or smooths the prototype associations of every candidate wrong answer. Since those associations enter both misconception-state updates and wrong-answer matching, changing τmis can also change candidate scores and rankings.
What τans changes
At fixed trained scores, a positive τans changes probability sharpness without changing candidate rankings. During model training, however, τans affects the loss gradients and may change the learned scores.