Causal Transformer: Toward Deductive Reasoning in Large Language Models

This work introduces the Causal Transformer (CT), a theoretical framework aimed at advancing the reasoning capabilities of large language models beyond purely statistical prediction. While current Transformer-based architectures excel at approximating conditional probabilities, they lack explicit representations of causality, logical consistency, and calibrated uncertainty. The proposed framework integrates four complementary components: latent-space causal inference, variational free-energy minimisation, category-theoretic logical constraints, and Bayesian-conformal uncertainty quantification. Together, these elements are designed to address key limitations of modern LLMs, including hallucinations, spurious correlations, and overconfident predictions. The contribution is intentionally twofold: on one hand, a practically applicable uncertainty module providing calibrated abstention mechanisms; on the other, a broader architectural vision outlining a path toward more robust, interpretable, and deductive AI systems. This work is not presented as a ready-to-deploy solution, but as a structured theoretical proposal for future research in causal and reasoning-aware language models.