zrmarine/geometry-of-reasoning

The Geometry of Reasoning

An interactive AR-style data visualisation for the paper "The Geometry of Reasoning" by Ryan Marinelli.

By segmenting and embedding internal reasoning steps, we track cosine distance between thoughts to detect when an LLM gets off track — without needing to parse the text itself.

Features

• —Trajectory tab — 3D vector space showing labelled thought nodes, coloured by cosine distance. Drag to orbit, scroll to zoom, inject logical steps from a categorised menu (deductive, inductive, backtracking, correction).
• —Model Comparison tab — Side-by-side visualisation of the three reasoning archetypes: the 1.7B Incrementalist (fine, consistent steps), the 7B Leaper (large scattered jumps), and the 14B Minimalist (sparse, levelling-off) — illustrating the paradox of scale.
• —Corrector tab — Animated pipeline showing real-time anomaly detection: cosine distance bars, threshold crossing, router activation, and fallback model correction.

Source

GitHub: rymarinelli/Distance-Between-Thoughts

Embedding model: all-MiniLM-L6-v2 → 384D vector space