Mengnan Du

Assistant Professor

md748@njit.edu

Research Areas: Trustworthy AI, explainability, fairness, natural language processing, large language models

Alignment of Multi-modal LLMs

Our research focuses on addressing the critical alignment issue in multi-modal Large Language Models (LLMs), with particular emphasis on VisionLanguage Models (VLMs) and Protein-Language Models (PLMs). We investigate fundamental challenges in achieving robust alignment between diverse data representations through three key perspectives: visual-linguistic alignment, proteinstructural alignment and cross-modal integration. Our work includes developing methods to bridge representation gaps between modalities, understanding alignment mechanisms in different architectures and improving model reliability through better training and evaluation approaches. By innovating in alignment techniques, evaluation frameworks and theoretical foundations, we aim to enhance the capability of LLMs to work seamlessly across different modalities while maintaining semantic consistency and factual accuracy.

Shuai Zhang

Assistant Professor

sz457@njit.edu

Research Areas:LLMs, In-context Learning, Transformer, Mamba, MOE, Feature learning, Generalization guarantees

Advancing Leaning Theory with Emerging LLM Structural Components

LLMs have demonstrated exceptional performance across many domains. This success is largely attributed to their “emergent” abilities, enabling them to solve complex, unseen tasks not explicitly anticipated during pre-training. A key example is in-context learning (ICL), where LLMs learn new tasks from just a few contextual examples, without fine-tuning. These capabilities are believed to stem from transformerbased pre-trained models, though the theoretical mechanisms remain unclear. Our goal is to advance the current feature-learning framework by analyzing the training dynamics of transformer-based LLMs and investigating the unique role of the attention mechanism. Additionally, we aim to extend this framework to incorporate emerging architectural components in LLMs, such as structured state space models and Mixture-of-Experts and work towards designing more efficient model architectures with strong generalization guarantees.

Guiling “Grace” Wang

Distinguished Professor

gwang@njit.edu

Research Areas: Deep Learning, AI in Finance, AI in Transportation, Large Language Model (LLM)

Enhancing Controllability and Explainability in Flowchart Understanding with LLMs

Flowchart understanding, often reliant on visionlanguage models (VLMs), faces challenges in controllability and explainability. Users have limited ability to influence processing beyond input modification and errors are difficult to trace due to opaque reasoning processes. To address these issues, we propose a two-stage framework: a VLM converts flowchart images into customizable text representations and an LLM performs reasoning and question-answering on the text. This approach enhances user control, isolates processing errors for improved explainability and promotes modularity by enabling integration with advanced reasoning tools. The framework’s structured intermediate representations also provide a foundation for generalizing to other multimodal tasks, improving usability and reasoning capabilities.

Advancing Creative Problem-Solving in Mathematics with LLMs

While research on Large Language Models (LLMs) has extensively explored their problem-solving capabilities, their potential for creativity in mathematical reasoning remains underexamined. To bridge this gap, we present CreativeMath, a framework designed to evaluate and enhance LLMs’ innovative reasoning abilities in mathematics. Published at AAAI 2025, CreativeMath introduces a benchmark comprising problems spanning middle school curricula to Olympic-level challenges, systematically assessing the creative problem-solving skills of LLMs. This study sheds light on both the strengths and limitations of LLMs in fostering mathematical creativity and offers a robust benchmark to advance our understanding of their cognitive potential.

Advancing Logical Reasoning in LLMs with FaultyMath

Large Language Models (LLMs) excel at solving standard mathematical problems but often fail to detect logical inconsistencies, raising questions about their ability to reason beyond rote calculation. To address this, we present FaultyMath, a benchmark for evaluating LLMs’ capacity to identify and reason about faulty math problems. FaultyMath encompasses diverse categories, including algebra and geometry, with varying difficulty levels and fault types such as contradictions and common-sense violations. It assesses LLMs’ performance in detecting flawed problems, incorporating hints and providing reasoned explanations. This research underscores the limitations of current LLMs in logical reasoning and establishes a foundation for enhancing their cognitive capabilities, fostering more robust and trustworthy AI systems.

Revolutionizing Table Question Answering with Secure and Efficient LLM Solutions

Table-based question answering with Large Language Models (LLMs) typically requires embedding entire tables into prompts. This approach faces challenges such as context window limitations, high computational costs and data leakage risks, particularly for large tables. To address these issues, we propose DataFrame QA, a task and framework that generates Pandas queries for information retrieval and analysis on tables. By using only table column names and data types, this approach ensures data privacy, reduces token usage and enhances efficiency, providing a foundation for secure and scalable LLM-powered tabular data analysis.