I’d like to introduce the latest iterations of our work: O2ARC 3.0 and ARCLE. O2ARC 3.0 is a competitive arena designed for engaging gameplay but also for advancing our work in reinforcement learning through ARCLE.

ARCLE is now ready to support the development of reinforcement learning agents, leveraging the actions and states you’re familiar with in the O2ARC environment. While the current actions are complete sets that solve ARC challenges by human standards, they are too low-level for training effective models. To address this, we’re contemplating integrating more comprehensive Domain-Specific Languages (DSLs), though they present substantial challenges. To build effective reinforcement learning agents on top of ARCLE, we need a strong community effort. I encourage everyone to explore O2ARC 3.0 and start working on models with ARCLE.


Code and application: