CauSim: Scaling Causal Reasoning with Increasingly Complex Causal Simulators

Researchers have introduced CauSim, a novel framework designed to address the persistent struggle of large language models (LLMs) with causal reasoning. While LLMs excel in mathematics and coding, they often fail in causal tasks due to the scarcity of ground-truth data and the complexity of non-executable causal systems. CauSim transforms this challenge into a scalable supervised learning problem by constructing increasingly complex causal simulators. These simulators are executable structural causal models (SCMs) incrementally built by LLMs, allowing for verifiable answers to causal queries. The framework operates across representations by formalizing non-executable causal knowledge into code for data augmentation and translating executable SCMs into natural language for supervision. The study demonstrates that CauSim enables generalization across representations, consistent performance gains through curriculum scaling and increased data volume, and LLM self-improvement via self-generated simulators. This approach effectively leverages existing domain knowledge to enhance causal reasoning capabilities in AI systems.