Written by Matthew Busekroos | Produced by Andrew Zucosky
Originally from Taiwan, PhD student Yung-Sung Chuang moved to the United States to embark on his studies at MIT CSAIL. Prior to MIT, Chuang completed his undergraduate degree at National Taiwan University. According to Chuang, MIT has world-class research groups, many of which he had already been following while in Taiwan. Chuang said the opportunities here are incomparable to what he had access to back in Taiwan, which is why he chose to come to the U.S. and pursue his PhD degree at MIT.
Chuang now works alongside Professor Jim Glass in the lab’s Spoken Language Systems Group.
“[Professor Glass] is a brilliant advisor,” Chuang said. “He is always supportive of my research ideas and projects, quickly grasping the challenges I face and offering valuable insights. [Prof. Glass] understands the importance and value of my research, trusts my decisions, and gives me a lot of freedom to pursue the directions I’m passionate about. He doesn’t interfere too much with my ideas, allowing me the space to think and explore independently.”
Chuang’s research focuses on NLP, specifically the factuality of large language models (LLMs). According to Chuang, these AI models, like the ones behind ChatGPT, have shown remarkable capabilities in assisting with everyday tasks. However, their tendency to hallucinate false information is a significant barrier to deploying them in high-risk situations, such as clinical or legal settings.
“I’m deeply curious about improving the factuality of LLMs by exploring their internal mechanisms,” Chuang said. “In a previous project, DoLa, we studied how LLMs store facts within the parameters of their transformer layers. We discovered that the early layers encode lower-level information like grammar, while the later layers store more factual or real-world knowledge. By enhancing the impact of these later layers through contrastive decoding, we improved the factuality of LLMs by 12-17% without needing to fine-tune the model.”
Chuang added that more recently, in the Lookback Lens project, the group aimed to understand why LLMs sometimes ignore context information and produce unsupported responses.
“We hypothesize that these ‘contextual hallucinations’ are closely tied to the attention mechanism within the transformer architecture,” he said. “To address this, we developed Lookback Lens, a lightweight classifier that detects and mitigates hallucinated outputs by analyzing the ratio of attention weights on context versus newly generated tokens. Lookback Lens successfully identifies hallucinations and reduces them by 9.6% in document summarization tasks, giving us better insights into the origins of these errors and how to mitigate them.”
Chuang said the ultimate goal of his research is to reduce or even eliminate hallucinations in LLM-generated text, which would greatly enhance the deployment of these models in real-world applications.
“Currently, AI models’ tendency to make mistakes and hallucinate limits their use in critical decision-making,” he said. “By minimizing these hallucinations, we can increase the trustworthiness and reliability of AI systems, enabling broader applications in vital areas like healthcare, legal advice, and education.”
Chuang added that curiosity is the main force behind his research.
“The concept of a general-purpose AI system is fascinating, yet the internal mechanisms that make it work are still not fully understood,” Chuang said. “This curiosity drives me to uncover the hidden processes that power AI systems.”
Upon finishing his work at CSAIL, Chuang said he plans to continue researching LLMs, ideally staying within research institutions, whether in academia or industry. He said his goal is to contribute to the development of LLMs, particularly in enhancing their accuracy and reliability.
You can find more information about Yung-Sung Chuang on his personal website: https://people.csail.mit.edu/yungsung/.