How Do Transformers Learn Variable Binding in Symbolic Programs?

A reference guide for key technical terms relevant to the background and design of the experiment.

Visualize variable assignment programs as interactive directed graphs.

Interactive heatmap showing model predictions across checkpoints and test programs. Click cells to copy programs for detailed analysis.

Track the model's performance across 105,400 training steps, revealing three distinct learning phases.

Analyze the predictions of individual training checkpoints across different attributes with interactive bar charts or Sankey diagrams.

Generate a report for each model checkpoint showing different metrics including accuracy by feature, prediction distributions, confusion matrices, and feature importance.

Compare multiple training checkpoints to visualize differences in accuracy, prediction patterns, and feature-based performance across training stages.

Visualize how the model's internal logit values evolve across layers and training steps. Compare logits for correct answers vs first line constants.

Visualize the results of causal intervention experiments on attention heads to understand how different heads affect the model's ability to process variable chains.

Visualize how variable names and numerical constants organize into distinct clusters in the model's representation space during training.