What question did this study set out to answer?

The aim is to examine how large language models demonstrate apparent causal reasoning and the impact of semantic scaffolding.

March 4, 2026Open Access

The Illusion of Causality in LLMs: A Developmentally Grounded Analysis of Semantic Scaffolding and Benchmark–Capability Mismatches

View Full Paper

Ask AI

Bookmark

Authors

This analysis demonstrates causal reasoning limitations in large language models, highlighting semantic influences on performance.

The aim is to examine how large language models demonstrate apparent causal reasoning and the impact of semantic scaffolding.
Analyzed benchmark evaluations of LLMs in causal reasoning tasks.
Utilized synthetic causal micro-worlds to test LLM responses under different label conditions.
Compared model performance using meaningful variable labels versus non-semantic coded labels.
LLMs selected the correct causal structure more consistently with meaningful labels.
Misidentification of causal models occurred frequently with non-semantic labels.
Divergences were evident in conditions requiring suppression of misleading associations.

Daisuke Akiba (2026) studied this question.

View Full Paper

Ask AI

Bookmark

March 4, 2026Open Access

View Full Paper

Ask AI

Bookmark

Authors

This analysis demonstrates causal reasoning limitations in large language models, highlighting semantic influences on performance.

The aim is to examine how large language models demonstrate apparent causal reasoning and the impact of semantic scaffolding.
Analyzed benchmark evaluations of LLMs in causal reasoning tasks.
Utilized synthetic causal micro-worlds to test LLM responses under different label conditions.
Compared model performance using meaningful variable labels versus non-semantic coded labels.
LLMs selected the correct causal structure more consistently with meaningful labels.
Misidentification of causal models occurred frequently with non-semantic labels.
Divergences were evident in conditions requiring suppression of misleading associations.

Daisuke Akiba (2026) studied this question.

View Full Paper

Ask AI

Bookmark