Conceptual networks reveal idioms cluster by meaning, not language
A new paper introduces an interpretable network-based framework for representing idiomatic expressions across eight languages. By annotating 160 idioms with binary conceptual features and computing pairwise similarities, the authors find that idioms cluster by conceptual schema rather than by language, offering a cognitively grounded approach to cross-linguistic idiom analysis.
Entities
Related
Neural Collapse Is Forbidden: Information Floors in Language Models
A new paper argues that within-class variance in language model representations is not incomplete neural collapse but allocated information storage obeying a law. Across 14 models, macro-category structure carries only 4-12% of representational variance, while within-token context carries 79-91%, stable across a 100x parameter range.
Blog post explores why LLMs produce predictable metaphors and how architecture might reduce attractor pull
A blog post titled 'Escaping the Attractor' examines why large language models tend to produce similar metaphors (e.g., 'Time is a River') when prompted, attributing this to attractors in the embedding space. The author suggests that architectural changes could make models less predictable, building on earlier ideas about shared embedding geometries across models.
PRecG: Legal Precedent Retrieval with Graph Neural Networks and Rhetorical Role Segmentation
A new paper introduces PRecG, a method for legal precedent retrieval that uses graph neural networks and rhetorical role segmentation to capture nuanced legal meanings. Unlike existing approaches that treat documents as monolithic texts, PRecG models the rhetorical structure of legal cases, improving retrieval accuracy by distinguishing the contextual significance of legal entities.
Anthropic's 'J-Space' research draws critique for missing observer effect
Anthropic published research on 'Verbalizable Representations' characterizing a 'J-Space' as a global workspace for reasoning. A Reddit commenter critiques the analysis for omitting the observer effect, citing cognitive science literature.
Paper derives exact dynamics of linear representation learning in neural networks
A new arXiv paper presents exact solutions for how linear concept representations emerge during neural network training, providing a mathematical framework for the dynamics of abstraction. This work formalizes the linear representation hypothesis, which underpins interpretability methods like linear probes and activation steering, and could guide future training and control techniques.