Implicit Causal Graph Construction in Text via Chain Discovery

A new study proposes using large language models to infer the hidden intermediate steps between a cause and its effect in text, moving beyond the standard practice of mapping only observable events in causal graphs [1]. The research, submitted to arXiv in April 2026, treats each described cause-effect pair as the start and end point of an underlying, latent causal graph and uses LLMs to fill in the missing links [1][2]. The work compares end-to-end graph construction with methods that frame the task as causal chain discovery, where graphs are built by aggregating inferred chains or by progressively expanding partial chains through an iterative search process [2]. The authors also explore "Wisdom of the Crowd" extensions that draw on causal knowledge from multiple LLMs, both in post-hoc aggregation and collaborative inference settings [2]. The validity of the inferred causal relations is evaluated against a manually curated database of 1,560 scientifically validated causal pairs [1][2]. The researchers propose this database-based evaluation as reliable, resource-efficient, and transferable to settings where ground-truth graphs are unavailable [2]. The concept of causality itself is a foundational abstraction indicating how the world progresses, and it is implicit in the structure of ordinary language as well as explicit in scientific notation [3]. The scientific method, which relies on hypothesis testing and experimental validation, has characterized science since at least the 17th century and provides the framework for validating such causal claims [5].