shared-language-causal-entanglement-structure
IN premise — summaries/2026/08/24/park-2023-linear-representation-s9-his-domain-was-vast-for-he-was-a.md
Created 2026-08-24T17:11:04+00:00
Language-translation concepts sharing a common language (e.g., English⇒French and French⇒German) show high causal inner product with each other, while non-overlapping pairs (German⇒Spanish) are nearly orthogonal, confirming geometric structure mirrors linguistic compositional relationships.
Summary
Translation-direction concepts that share a common language (like English-to-French and French-to-German both involving French) line up closely in the model's internal geometry, while pairs with no overlap sit at right angles to each other. This means the spatial structure naturally encodes which languages compose together, so downstream reasoning can exploit those relationships without needing to track them explicitly.