Geometric Interpretability: Understanding Neural Networks through Boundary Analysis
A framework for understanding neural networks through spatial structure rather than semantic content. We propose Geometric Boundary Analysis (GBA) for detecting knowledge boundaries in activation space, achieving 96.5% hallucination detection with perfect linear boundaries (0.3% curvature deviation, 1.000 symmetry, 148× concentration). Our grokking experiments reveal that class separation correlates more strongly with generalization (r=0.910) than boundary sharpness (r=-0.651), suggesting separation—not sharpness—may be key to generalization. The method requires only 30-second setup, <1ms inference overhead, and $0.50 cost with no model retra
Resource information
- Format
- zip
- Original repository
- 17329870
- DOI
- 10.5281/zenodo.17329870
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
A framework for understanding neural networks through spatial structure rather than semantic content. We propose Geometric Boundary Analysis (GBA) for detecting knowledge boundaries in activation space, achieving 96.5% hallucination detection with perfect linear boundaries (0.3% curvature deviation, 1.000 symmetry, 148× concentration). Our grokking experiments reveal that class separation correlates more strongly with generalization (r=0.910) than boundary sharpness (r=-0.651), suggesting separation—not sharpness—may be key to generalization. The method requires only 30-second setup, <1ms inference overhead, and $0.50 cost with no model retraining.