Large Causal Models from Large Language Models (Dec 2025)

Abstract

• Title: Large Causal Models from Large Language Models (Dec 2025)
• Link: http://arxiv.org/abs/2512.07796v1
• Date: December 2025

Summary:

This paper introduces DEMOCRITUS, a new paradigm for building Large Causal Models (LCMs) by extracting and organizing the vast causal knowledge latent in Large Language Models (LLMs). Unlike traditional causal discovery that relies on numerical data from narrow domains, DEMOCRITUS aggregates millions of specific causal claims from diverse fields into a coherent geometric structure using a pipeline that includes topic expansion, causal triple extraction, and a Geometric Transformer. The authors demonstrate the system’s ability to create navigable causal manifolds in domains ranging from economics and biology to the archaeology of the Indus Valley Civilization collapse, organizing fragmented text into structured Topos Causal Models.

Key Topics:

• Large Causal Models (LCMs)
• Large Language Models (LLMs)
• Causal Discovery
• Geometric Transformer
• DEMOCRITUS System
• Topos Causal Models
• Manifold Learning