"Human-Like Reasoning"... GIST Develops AI That Estimates 'Intentions' in Problem-Solving Processes

(Photo from left) Professor Kim Seondong of GIST AI Convergence Department, Postdoctoral Researcher Kim Sejin, Master's Graduate Hwang Sanha, Master's Student Lee Seungpil, Bachelor Graduate Lee Hoseong of Electrical, Electronics and Computer Engineering Department.

The Gwangju Institute of Science and Technology (GIST) announced on the 25th that the research team led by Professor Kim Seondong of the Department of AI Convergence has proposed a data augmentation technique that combines a learning algorithm for estimating and aligning the "intentions" embedded in human problem-solving processes with an AI generative model, thereby generating diverse problem-solving trajectories similar to those of humans.

The research team demonstrated that by combining these two approaches, artificial intelligence can go beyond simply producing correct answers and actually implement "human-like reasoning abilities."

While artificial intelligence excels at quickly deriving correct answers to given problems, it still lacks the reasoning ability to solve problems through step-by-step thinking processes like humans do. Humans experience trial and error when solving problems and attempt to reach the same goal through various methods. The accumulated problem-solving data from this process contains not just a sequence of actions, but also intentions. Learning these intentions is key to developing AI with "human-like reasoning abilities."

From this perspective, the research team developed an algorithm to estimate and align intentions within problem-solving processes and proposed a data augmentation technique using GFlowNet, one of the generative models, to create diverse problem-solving trajectories.

The former approach reflects human thinking processes, while the latter expands problem-solving trajectories to enhance generalization performance. These two approaches complement each other. Through this, the goal was to develop AI models with human-like thinking and generalization abilities, going beyond simply producing correct answers.

Various problem-solving trajectories and examples of intentions.

The research team analyzed human problem-solving data and classified cases where optimal solutions were not reached into three types: lack of functionality, inefficient attempts, and incorrect strategies, and incorporated these into the learning process. They then presented an algorithm to divide problem-solving trajectories into multiple steps, estimate and align the intention at each step, and incorporated this into AI learning to implement learning that mimics human thinking processes. In addition, by using the GFlowNet-based data augmentation technique to generate diverse problem-solving processes similar to those of humans, they significantly expanded the diversity and generalization performance of the training data.

GFlowNet generates multiple solution paths by probabilistically exploring various intermediate states. For the same input, it creates multiple trajectories using various combinations of operations such as rotation and symmetry, which are then used for training. This enables the model to learn a diverse distribution of strategies without being biased toward a particular problem-solving approach.

Problem-solving trajectory generation process based on GFlowNet.

The research team collected problem-solving trajectory data from actual people and supplemented insufficient trajectories by generating them with GFlowNet. As a result of applying this diverse problem-solving data to training, accuracy improved by approximately 5.85 percentage points compared to existing models (from 83.59% to 89.44%). This demonstrates that AI can think and generalize like humans.

Professor Kim Seondong said, "Humans often find answers using familiar shortcuts rather than always following the standard solution process, but in AI model training, it is common to use human-collected data without much consideration. By systematically applying the pre- and post-processing procedures presented in this paper, it will be possible to overcome the inherent limitations of data and develop AI that exhibits more desirable behaviors."

Dr. Kim Sejin stated, "This research is the first step toward implementing AI that not only gets the right answer, but also understands human problem-solving intentions and can learn diverse problem-solving processes on its own. In the future, it could be expanded to creative problem-solving AI, educational collaboration AI, and AI that adapts to new environments."

This content was produced with the assistance of AI translation services.