Constraint-Induced Invariant Semantics: A Unified Framework Bridging Weight Geometry, Dynamics, and Model Alignment
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
We propose a unified framework in which semantics is not a property of representations, but an invariant structure emerging from constraint-driven dynamics. Formalizing the constraint induced by modelparameters as the Fisher information operator, we show that weight-space geometry defines a Riemannian structure whose geodesic flow governs depth-wise representational dynamics. We prove that systemswhose constraint operators are close in Frobenius norm converge to attractors that are proportionallyclose, independent of weight-level differences. Empirical results across ten model pairs—spanning GloVe,Word2Vec, FastText, BERT-base, ALBERT, RoBERTa, and DistilBERT—show that constraint similarity (CKA on diagonal empirical Fisher matrices) consistently outperforms weight similarity in predictingsemantic alignment (ρ = 0.71 vs. ρ = 0.09, p < 0.01). These results establish a substrate-independenttheory of semantics and provide a new perspective on model interoperability and cross-architecture alignment.
Fields of Study & Subjects
Topic Modeling
Domain Adaptation and Few-Shot Learning
Explainable Artificial Intelligence (XAI)
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 30 March 2026
- Language
- EN
- Document type
- Preprint