In the fast-paced realm of open-world machine learning, Generalized Category Discovery (GCD) has emerged as a crucial task for identifying new classes within ever-evolving datasets. With the rise of multimodal data that includes text, images, audio, and video, traditional GCD methods, which often rely on parametric classifiers and single-modality inputs, face significant limitations. These approaches can lead to overfitting and hinder the ability to generalize to new categories effectively. This paper highlights the pressing need for innovative strategies that harness the richness of multimodal data to enhance contextual understanding and facilitate real-time category identification. We aim to establish a foundational framework for future GCD research, promoting a more agile and resilient approach to data classification in today’s complex information landscape. To achieve this, we propose a dynamic framework that integrates heterogeneous multi-agent systems, combining Large Language Models (LLMs) with diverse non-LLM methodologies. This approach not only enhances the adaptability and robustness of GCD solutions but also opens up transformative possibilities across critical fields such as autonomous driving, medical diagnostics, and social media analysis.

Large language models (LLMs) have been shown to struggle with complex logical reasoning tasks due to the inherent ambiguity and complexity of natural language. These challenges are further amplified when processing large and diverse datasets, increasing the likelihood of unfaithful reasoning and predictive hallucinations. However, LLMs can provide accurate responses when queries are clear and direct. Symbolic logic provides precise, well-defined rules that can help overcome ambiguity and support reasoning. In this work, we leverage symbolic logic’s precision to enhance LLMs’ logical reasoning capabilities by introducing the Graph of Logic (GoL) framework. GoL combines the power of graph structures with the strengths of LLMs and symbolic logic. GoL utilizes the precise rules of symbolic logic to infer new facts and detect LLM hallucinations effectively on complex datasets. Furthermore, GoL utilizes graph structures to support scalability for large datasets and tackle long dependencies, enabling efficient handling of complex reasoning tasks. We conduct extensive experiments across seven benchmark datasets, encompassing various types of reasoning. These include deductive, inductive, and abductive reasoning, each testing distinct aspects of logical inference. The experimental results demonstrate GoL’s advantage in improving the reasoning capabilities of LLMs. GoL outperforms the baselines with an average margin of 18.18% for the GPT-3.5 and GPT-4 models, outperforming the baselines for all datasets for the GPT-3.5 model and six out of seven datasets for the GPT-4 model1.

Logical reasoning is crucial in knowledge graphs, enabling the discovery of latent knowledge and facilitating various downstream tasks such as question-answering and knowledge discovery. Despite the promising results of state-of-the-art methods, their limited generalization capabilities in handling out-of-distribution (OOD) queries and KG data pose a significant challenge. This challenge arises from the complex variations in logical queries and the dynamic nature of graphs, resulting in two OOD scenarios at the query and KG data levels. To address this challenge, we propose a novel, generic framework that handles bi-level OOD logical reasoning on KGs by jointly modeling OOD queries and OOD KG data. First, we start with the uncertainty representation module to handle the KG-level OOD. Then, we further design an adversarial learning module to efficiently improve reasoning over query-level OOD. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework in handling these two levels of OOD scenarios.