Research

Research

Overview

TokenPrint is not just an educational app; it is a tool for mechanistic interpretability. This section outlines the philosophy behind TokenPrint's design, how it compares to existing tools, and the academic foundations it builds upon.

Why it matters

If a visualizer abstracts away the complexity of a neural network by inventing metaphors (like "the model imagines a dog"), it fails as a scientific tool. TokenPrint's design is strictly grounded in the reality of linear algebra and tensor mechanics.

Section Contents

• Related Papers: The academic literature that informed TokenPrint's implementation (e.g., RoPE, SwiGLU, Mechanistic Interpretability).
• Existing Visualization Tools: A comparison between TokenPrint and tools like Netron, TensorBoard, and LLM Visualization by bbycroft.
• Design Philosophy: The core tenets guiding TokenPrint: "Real data over decorative simulation."
• Scientific Accuracy: How TokenPrint mechanically guarantees that its geometry and outputs are not fabricated.

Guiding Principles

1. No Math.random(): If a number is displayed, it was generated by a forward pass or parsed from a model configuration.
2. Architecture-Aware: An MLP block is not a generic box. If the model is Llama, it must be rendered as SwiGLU.
3. Data-Driven Geometry: A 32-head model must explicitly render 32 physical attention blades in the 3D scene.

Related pages

• Project Roadmap
• Design Philosophy

Further reading

• Verification Docs

Navigation

Previous	Home	Next
Tensor Format	Home	Related Papers