park2025-validation-models-concepts

Summary

This confirms that the geometric structure findings from Park et al. (2025) were not just a one-off result on a single model, but held up across two major model families (Gemma and LLaMA-3) and on a substantial, hierarchically structured concept set drawn from WordNet. It gives downstream reasoning in the system a solid empirical anchor when relying on those geometric claims, since they rest on broad validation rather than a narrow test.