Shape Happens: More LLM Geometry w/ Federico Tiblias

Language models don't just store facts, they mold them into geometry: Calendar dates trace circles. Time durations compress logarithmically. Seasons cluster. Federico Tiblias (TU Darmstadt) introduces Supervised Multi-Dimensional Scaling (SMDS), a method for testing what shape a concept takes inside a model's representations and shows these shapes are consistent across model families, reshape dynamically depending on the task, and actively support reasoning. Perturb the geometry and performance degrades; preserve it and the model thinks clearly. A new lens on what it means for a model to "understand" something.