A new AI framework tries to fix a quiet flaw in how language models build causal graphs: they tend to leave real connections out.
Researchers built MaSCoD, a multi-agent system that gathers candidate third variables and structural patterns before an LLM judges whether a direct causal edge exists between two things. They tested it on three datasets, Auto-MPG, DWD, and Sachs, using GPT-5.4 as the main model and GPT-4o to check the results held up. Across all six dataset-and-model combinations, the version that organizes structural hypotheses first ("Full") beat the version that builds them during judgment itself, scoring higher on both recall and F1. That gain came with a cost: more false positives, and the extra accuracy on known reference edges only showed up in two of the six test combinations.
Causal discovery tools already lean toward missing real relationships rather than inventing fake ones, which is arguably the worse failure mode if you are using this to guide a scientific or business decision. Pre-organizing the context before asking a model to judge cause-and-effect looks like a real lever for reducing that blind spot. But the paper is upfront that the effect is dataset- and model-dependent, not a universal fix, and that fixing omissions by giving the model more to work with just trades one error type for another.
It is a sensible next question for anyone building causal-inference pipelines on top of LLMs, but it is a lab result on three modest datasets, not a benchmark win worth quoting in a pitch deck.