Multi agent system guides causal graph generation to reduce missed links
MaSCoD: A Multi-Agent Framework for Structural-Context-Guided Candidate Causal Graph Generation
Artificial Intelligence
Summary
Finding cause-and-effect relationships from data is tricky, especially when important connections might be missed early on. The authors present MaSCoD, a system that uses multiple agents to organize possible factors and patterns before deciding if one factor directly causes another. They tested MaSCoD with different datasets and versions of large language models, showing it helps keep track of more true connections but sometimes adds some incorrect ones. Their work suggests that planning the structure before making decisions helps reduce missing important cause links.
What this means in practice
- •For data science teams: Improve causal relationship detection from data by organizing candidate causes before final evaluation in analysis pipelines.
- •For business intelligence teams: Enhance insights by reducing prematurely dropped potential causal links when modeling relationships in complex datasets.
Authors
Yudai Nakada, Yuichiro Nishiura, Jin Michael Splichal
Abstract
Large language models (LLMs) have been applied to causal discovery, but candidate-graph generation rarely treats premature omission of potentially relevant causal relations as an explicit design objective. We propose MaSCoD, a multi-agent framework that organizes candidate third variables and local structural patterns before direct-edge judgment. We evaluate MaSCoD on Auto-MPG, DWD, and Sachs using GPT-5.4 as the primary backbone and GPT-4o for replication. MaSCoD exhibits a dataset- and backbone-dependent retention-selectivity profile rather than uniform superiority. Across all six dataset-backbone settings, Full, which supplies structural hypotheses before direct-edge judgment, achieved higher mean Recall and F1 than No Phase 1, which instead constructs them within the judgment procedure, while also increasing false-positive rates. Additional reference-edge retention over all evaluated baselines was observed on DWD with GPT-5.4 and on Sachs with GPT-4o, rather than uniformly across settings. Partial ablations showed that supplying both information components did not always outperform supplying only one. For GPT-5.4, stage-wise analysis showed that the Full-No Phase 1 retention gap was already present after direct-edge judgment, while reconciliation introduced additional reference-edge loss for Full on Sachs. These findings support structural pre-organization as an explicit design and evaluation target for omission control and motivate evaluating context construction jointly with its utilization in judgment.