Reproduce “Abductive Logical Rule Induction by Bridging Inductive Logic Programming and Multimodal Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifei Peng, Yaoli Liu, Enbo Xia, Yu Jin, Wang-Zhou Dai, Zhong Ren, Yao-Xiang Ding, Kun Zhou
ILP-CoT, a neurosymbolic method from Zhejiang and Nanjing Universities, enhances abductive logical rule induction by integrating Multimodal Large Language Models (MLLMs) for initial symbolic grounding and hypothesis generation with Inductive Logic Programming (ILP) for formal verification. The approach improves logical reasoning accuracy on benchmarks like CLEVR-Hans (81.85% test accuracy) and ARC, effectively rectifying MLLM hallucinations and enabling interpretable rule learning for tasks such as text-to-image customization.
Key insights
- MLLMs can effectively propose syntactically correct rule structures (meta-rules) even when their semantic factual grounding contains hallucinations, providing a crucial inductive bias for ILP.
- Automatically transforming these MLLM-generated rule structures into meta-rules significantly prunes the ILP system's hypothesis search space, making abductive logical rule induction tractable from unstructured inputs without human expertise.
- Formal verification by the ILP module, coupled with a failure reflection loop, effectively rectifies MLLM-generated perceptual hallucinations, leading to robust and interpretable logical rules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.