Exploring Reasoning with Flow Matching on the ARC-AGI Benchmark
Ma Kai is a talented Master’s student from Sorbonne Université, France, who has joined University Metropolitan Tirana, as a research intern for 5 months. During his internship, Ma Kai will work closely with UMT’s research team, contributing his expertise and fresh perspective to advance the ongoing research efforts in the field of reasoning and Artificial General Intelligence (AGI).
His collaboration with UMT’s researchers will not only strengthen the project’s academic depth but also foster a dynamic exchange of ideas, enhancing the research topic through innovative approaches and rigorous analysis. This partnership reflects UMT’s commitment to nurturing international collaboration and cultivating impactful research outcomes.
Context and Background
In 2019, François Chollet, creator of Keras, the widely adopted open-source deep learning library, introduced the seminal paper On the Measure of Intelligence. In it, he proposed the Abstract and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) benchmark, a dataset designed to assess a system’s ability to generalize and learn efficiently on previously unseen tasks.
Chollet’s view of intelligence centers on the efficiency of skill acquisition: the faster and more effectively a system can learn to solve new problems, the more intelligent it can be considered.
The ARC-AGI benchmark challenges AI systems to demonstrate more than pattern recognition; it requires them to reason, generalize knowledge to novel scenarios, and compose learned skills in innovative ways, hallmarks of true intelligence.
Recently, flow matching algorithms, which have already achieved remarkable success in areas like image and video generation—have emerged as a promising framework to address such reasoning challenges, due to their adaptability and compositional properties.
Internship Objective
During his internship at UMT, Ma Kai will:
- Explore reasoning strategies based on flow matching methods.
- Apply these strategies to the ARC-AGI benchmark to evaluate their effectiveness.
- Investigate how flow matching can enable models to perform structured reasoning and tackle diverse problem-solving tasks in a more compositional and adaptable manner.
This research sits at the intersection of reasoning, generalization, and machine learning, pushing the boundaries of what current AI models can achieve in terms of Artificial General Intelligence.
