Beijing, China — Researchers from Beijing Institute of Technology and other leading institutions have developed a novel approach for improving Multi-Hop Question Answering (MHQA) tasks, which require models to reason over multiple relevant facts to answer complex questions. The new method, called CausalBridgeQA, integrates causal inference into the MHQA process to address persistent issues of reasoning breakdowns and feature spurious correlations, which are common challenges in current models.

To solve the problems, the team, led by Xu Jiang, published their new research on 15 March 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature. Their research shows that by incorporating causal relationships into the question answering process, their method significantly improves the accuracy and robustness of MHQA tasks, even when dealing with ambiguous or complex reasoning chains.

“MHQA tasks require deep reasoning, where the model must connect facts across different paragraphs and identify relationships between them. Traditional models often fail because they rely on single facts or are misled by irrelevant data,” said Xu Jiang, a PhD student at Beijing Institute of Technology. "Our approach allows the model to extract causal relationships, which strengthen the reasoning process and lead to more accurate results."

The CausalBridgeQA method works by first extracting causal relationships from the context and then transforming the original question into one enriched with this causal information. This enhanced question is then used in a modified MHQA system, which incorporates a knowledge compensation mechanism to deal with difficult or frequently misanswered questions.

The team validated their method with a series of experiments on popular datasets like HotpotQA, SQuAD, and TQA. The results demonstrated that CausalBridgeQA outperforms existing models, achieving higher accuracy and F1 scores in multi-hop question answering tasks. Notably, it reduces the common problem of reasoning breakdowns and ensures the integrity of reasoning chains, even in the presence of misleading or irrelevant information.

“The ability to identify and integrate causal relationships not only enhances the model’s performance but also makes the reasoning process more transparent and interpretable,” added Yurong Cheng, an associate professor at Beijing Institute of Technology.

Looking forward, the team plans to expand the method’s capabilities by improving causal relationship extraction and refining the knowledge compensation mechanism to address even more complex question answering scenarios. The integration of causal reasoning could lead to significant advancements in artificial intelligence applications, particularly in areas requiring deep understanding and complex decision-making.