Abstract
Large language models (LLMs) have rapidly developed, and their reasoning capabilities have become a hot research topic. However, there is still limited exploration of abductive reasoning.
The multi-perspective and multi-level causes are one of the core challenges of abductive reasoning, which cannot be effectively solved by existing methods. We construct a specialized dataset named DeepAbduction, designed for tracing the causes of pollution and disease, addressing the lack of datasets in this field.
We propose the Inverse-Forward Abductive Reasoning (IFAR) framework for multi-perspective and multi-level abductive reasoning with LLMs. IFAR is zero-shot and combines generalized backward reasoning with relation-by-relation forward verification. Experimental results show that IFAR achieves an approximately 40% improvement in F1 score compared to other methods under mainstream LLMs, while maintaining a balance between recall and precision.
Furthermore, IFAR enhances the performance of non-reasoning LLMs to surpass those trained for reasoning, and remains effective when applied to the latter. Code will be released after acceptance of our work.
Blogger's Review: The introduction of the IFAR framework offers a novel approach to tackling key challenges in causal reasoning, particularly in dataset construction and enhancing reasoning capabilities. Future research could explore its application potential in other domains.