MIT PhD students interning at the MIT-IBM Watson AI Lab are researching ways to improve the efficiency and accuracy of AI systems in understanding and communicating through natural language. The team, including Athul Paul Jacob, Maohao Shen, Victor Butoi, and Andi Peng, aims to enhance each stage of the process involving natural language models, from comprehending and trusting words and their context to applying the information to real-world scenarios.
Jacob, who uses game theory to optimise language model outputs, is grounded in the interactive board game “Diplomacy.” His team came up with an algorithm that motivates an AI system to provide more truthful and reliable answers. Shen’s project targets on correcting miscalibration of the AI’s confidence in its generated answer. The technique allows the AI to generate a free text, converted into a multiple-choice task, which would tell if the model is overconfident or underconfident.
Conversely, Butoi and his team improve AI models by enabling vision-language models to reason compositionally and understand key phrases. They have developed a technique called low-rank adaptation of large language models (LoRA) that allows the AI model to comprehend complex coordinates such as ‘left’ and ‘right.’
In the realm of robotics, Andi Peng and his team are designing AI models to help individuals with physical boundaries. They believe that AI systems, although designed for autonomous tasks, should help people in a manner that is understandable to them. The group is constructing two embodied AI models in a virtual world to help individuals with physical limitations. They focus on strengthening the model’s decision-making abilities, two-way communication, understanding of a physical scene, and how to contribute best.
