Relation Geometry in Semantic Space of Language Models

Reversing Arrows in Large Language Models

This work presents the first systematic study of inverse relation directionality in LLMs, using a benchmark consisting of 5,457 instances spanning 27 distinct inverse relation labels and reveals systematic asymmetries in inverse relation classification across LLMs.

From Word Embeddings to Semantic Projections: Interpretability and Context in Web-Scale Semantic Analysis

Recent developments in NLP and web-scale document analysis have increasingly emphasized the importance of interpretability and contextual dependence in semantic representations. Although modern word embeddings achieve remarkable empirical performance, their semantic structure is often difficult to interpret, since meaning is encoded through latent geometric relations in high-dimensional spaces. This paper discusses an alternative conceptual framework based on explicit contextual semantic relations. Building on ideas from distributional semantics, co-occurrence analysis, and fuzzy set theory, the study revisits semantic projections and related count-based representations as interpretable directional semantic structures for semantic analysis in document corpora and web-based information environments. In this setting, several classical association measures, including PMI and related transformations, may be understood as derived from simpler conditional semantic projections. The methodology is illustrated through a comparative analysis of semantic associations related to “ChatGPT” across general web-scale data and specialized scientific repositories. Our results demonstrate that semantic projections effectively capture persistent contextual structures while remaining sensitive to corpus-specific discourse communities. The resulting perspective emphasizes interpretability, asymmetry, contextual dependence, and direct empirical meaning as central principles for semantic representation.

From Unified Correspondence to Parametric Correspondence

Local and Global Regimes of Geometric Complexity in Language Model Representations

A scale-dependent transition between two ID regimes is found: at low lexical diversity, conditions with fewer unique final words produce higher ID, while at high lexical diversity, this ordering reverses, and conditions with more unique words produce higher ID.

Similarity All The Way Up: Multilingual Generalization in LLMs Relies on Language-Level Similarity Structures

As Large Language Models (LLMs) grow more capable across diverse tasks, their (in)ability to generalize remains difficult to quantify and poorly understood beyond limited domains. In particular, LLMs are known to struggle generalizing multilingually, to languages outside of English, and that are poorly attested in their training data. To understand why this may be, and what enables some models to perform better than others, we turn to a long history of work across the cognitive sciences, arguing that successful generalization derives from appropriate representations in similarity space. We look at how well LLMs'representations capture the hierarchical similarity structure between distinct languages. Strikingly, we show LLMs'latent representations largely recover the hierarchical structure of the Indo-European language family tree -- grouping languages that are members of the same subfamily closely together in representation space. Furthermore, we show that the degree to which models reflect the similarity structure of languages correlates with their performance on XNLI, a multilingual natural language inference benchmark. This extends classic work on similarity-driven generalization at scale, showing how models that represent similar languages similarly generalize better from one language to another.

: An Open Interface for the Study of Multilingual Discourse Relations