Visualizing Model Weight Space: 3D Heatmaps Reveal Fine-tuning Impacts on Accuracy

CatAstro_Piyush · x · 2026-08-14

The Neural Thickets blog released an interactive visualization called "Accuracy Landscape." By perturbing models around their pretrained weights, the project demonstrates how task accuracy fluctuates across different benchmarks.

The page features both 2D heatmaps (using RGB to map tasks like GSM8K, OlyBench, and Countdown) and interactive 3D surfaces. Users can directly rotate and zoom in the browser using OrbitControls to explore the performance landscape of models (e.g., 0.5B parameters) under weight perturbations.

Original post →

More from Research

Research channel →