Our paper Addressing and Visualizing Misalignments in Human Task-Solving Trajectories is accepted to KDD 2025 🎉

This work was conducted in collaboration with our postdoc Sejin Kim and undergrad alum Hosung Lee, congrats on another ARC-AGI paper!

In this paper, we traced users’ task-solving behaviors and analyzed how they deviated from the optimal problem-solving strategy. We found that, even when users correctly identified the underlying analogy of a task, they often relied on shortcuts or engaged in repetitive actions—reflecting inherent human biases, especially when the problem-solving interface is less than ideal. (Note: O2ARC is nearly perfect—we spent three years refining it.)

Although the system logs reveal how users cleverly leveraged available features or found workarounds, a central challenge—particularly in the context of Inverse Reinforcement Learning (IRL)—is determining how accurately one can infer users’ original intentions from such behavior. In other words, to what extent can we recover authentic problem-solving intent from actions that are shaped, or even distorted, by interface constraints and cognitive biases?

We have another good news, our paper on GFlowNet Trajectory Augmentation has been accepted to Transactions on Machine Learning Research (TMLR)!

Congratulations to all the authors! This project originated from Sanha Hwang’s master’s thesis and was further developed together with Seungpil Lee’s hard work on LLM experiments under Sejin Kim’s guidance, leading to its successful acceptance. I’m truly glad about the publication on Sanha’s master thesis, and I hope this work will be meaningful and valuable to those who believe the ARC-AGI challenge is about learning to solve problems through step-by-step reasoning. hashtag#TMLR hashtag#ARC hashtag#GIST hashtag#AI

#GIST #ARC #KDD #AI #CogSci #TMLR