Towards Logical Reasoning and Learning in Open and Dynamic Environments

Abstract

We live in an increasingly open and dynamic world where knowledge is constantly evolving, and Artificial Intelligence (AI) systems must adapt to newly added information. A crucial aspect of AI systems is performing robust logical reasoning that makes reliable inferences, generates hypotheses, and extracts meaningful insights from vast and complex data. Logical reasoning is fundamental in high-impact applications such as medical diagnosis, autonomous driving, and scientific discovery. However, traditional AI models, designed under static assumptions, struggle with reasoning in open and dynamic environments. They fail to generalize beyond their training data, leading to unreliable conclusions when encountering novel entities, relationships, unforeseen scenarios, or incomplete knowledge. This limitation poses a significant barrier to these high-impact applications. For example, in biomedical research, knowledge graphs often contain thousands of entities and relationships representing gene-disease-compound interactions. While many of these relationships are well-established, new ones emerge as scientific knowledge evolves. This highlights the need for robust logical reasoning in open environments to handle distributional shifts, complex logical inferences, and knowledge discovery. As a result, a critical research question arises: "How can AI systems effectively reason in dynamic environments where knowledge is constantly evolving and uncertainty is inherent?" To tackle this challenge, this thesis introduces a comprehensive framework for enhancing logical reasoning in open and dynamic environments. The framework is structured around three core components (Aim1), Out-of-Distribution (OOD) Logical Reasoning, which develops techniques to enable AI models to generalize beyond their training data, especially in knowledge graphs where new queries and entities frequently arise. This involves characterizing uncertainty and distributional shifts, thereby identifying novel, incomplete, and uncertain data for logical reasoning tasks, ultimately enhancing the robustness of logical reasoning models. (Aim2) Graph-Augmented Logical Reasoning integrates symbolic logic with graph-based representations to enhance LLM's reasoning accuracy, interpretability, and robustness, addressing hallucinations and ambiguous inference. (Aim3) Applications: This aim highlights the real-world applications of open and dynamic logical reasoning, including Otrouha—an automated system for knowledge discovery in Arabic Electronic Theses and Dissertations (ETDs)—and a multi-agent framework for dynamic category discovery. It also includes a multi-agent framework for scientific hypothesis generation that integrates semantic processing and symbolic reasoning to operate over structured metadata, enabling the discovery of emerging research directions in evolving domains without requiring access to full text. By addressing the pressing challenges of evolving data and the inherent uncertainty in AI in these three aims, our framework unites complementary approaches that collectively drive robust logical reasoning. The Characterization of uncertainty and OOD in (Aim1) provides insights for(Aim2) to manage complexity, ambiguity, and unseen logical tasks. Moreover, (Aim2) utilizes symbolic logic, graph-based methods, and prompt engineering to demonstrate that these strategies can handle challenging and unseen complex logical reasoning. Building on this knowledge from (Aim1) and (Aim2), the proposed novel approaches can be deployed in real-world applications in (Aim3), yielding tangible impact. Ultimately, the synergy among these three aims forms a unified framework that robustly advances logical reasoning in open and dynamic environments, enabling AI to adapt, generalize, and support real-world knowledge discovery and decision-making.