Meet CBI fellow Xiang Yue

Meet Xiang Yue, one of the newest CBI fellows.

Tell us about yourself.

My name is Xiang Yue, and I’m honored to be a Carnegie Bosch Postdoctoral Fellow at Carnegie Mellon University. I received my Ph.D. from The Ohio State University, where I specialized in Natural Language Processing (NLP), particularly large language models (LLMs). My research focuses on enhancing the core reasoning capabilities of LLMs, allowing them to better understand, process, and generate human language. I’m particularly interested in how these models can be advanced to solve complex real-world problems and evaluated rigorously through real-world benchmarks.

What are your long-term goals?

In the long term, I aim to advance the field of NLP by improving the fundamental reasoning abilities of AI models. While LLMs have made significant progress in language understanding and generation, their reasoning capabilities—how they draw inferences, integrate information, and provide accurate responses—still have room for improvement. Enhancing these capabilities will be crucial to unlocking the full potential of LLMs in real-world applications, from healthcare to scientific discovery.

A key part of my vision is to develop techniques that enable LLMs to reason more like humans, handling tasks that require not just pattern recognition but also deeper, logical problem-solving. As I progress in my academic career, I hope to lead a research group focused on advancing these core technologies and applying them in high-impact areas. Additionally, I am committed to improving evaluation techniques for LLMs to better measure their performance on these reasoning tasks. My recent work on the multimodal LLM evaluation benchmark (MMMU) is one example of such efforts, aiming to ensure LLMs are rigorously tested on tasks that go beyond traditional benchmarks.

How will the CBI Postdoc Fellows program help you achieve these goals?

The CBI Postdoc Fellows program offers the perfect environment to pursue these goals. Carnegie Mellon University is home to a world-class research community with a strong focus on NLP and AI, and the collaborative opportunities across related fields are unparalleled. The fellowship’s interdisciplinary nature, along with partnerships like Bosch the connection with Bosch, will allow me to bridge my research with practical, real-world applications.

The resources and flexibility provided by the program will enable me to experiment with innovative techniques for enhancing the reasoning abilities of LLMs. I’ll also be able to expand on my work in evaluation—developing more rigorous benchmarks like the MMMU, which assesses multimodal models on their reasoning across text, images, and other data forms. These evaluations are crucial for understanding how well LLMs handle complex real-world scenarios.

What is the biggest challenge you hope to tackle?

One of the biggest challenges I aim to address is improving the reasoning capabilities of LLMs in real-world settings. Current models can produce impressive results, but they often struggle with tasks that require abstract thinking, logical reasoning, and handling unseen situations that go beyond their training data. Understanding how LLMs develop reasoning abilities, and where these capabilities originate, is a key question I hope to explore.

Another significant challenge is integrating knowledge from different modalities—such as text, images, and beyond—into LLMs’ reasoning processes. This is essential for advancing models capable of handling complex, multimodal tasks in real-world applications, but remains difficult for even the most advanced systems.

What obstacles do you see in addressing these challenges?

Several obstacles exist in improving LLM reasoning. One major challenge is the complexity of training models to go beyond surface-level language generation and engage in deeper, more abstract reasoning. Creating models that can generalize well to unseen tasks and develop reasoning skills independently of extensive human supervision is a difficult technical problem.

Another challenge lies in building models that effectively integrate multimodal data into their reasoning processes. Merging information from various sources to produce coherent, logical outputs requires both technical innovation and more refined evaluation metrics. 

Lastly, while my primary focus is on improving LLM reasoning, I also acknowledge the importance of ensuring that these systems remain safe and reliable as they evolve. The responsible deployment of LLMs in real-world settings, particularly in sensitive areas like healthcare, must be considered alongside any technical advancements.

Check back soon to meet the other CBI fellows and their faculty hosts!