Case Study: 2K Games achieves 80% ARC AGI 2 training puzzle solves with Strong Compute

2K Games reaches 80% ARC puzzle solve rate with Strong Compute

Strong Compute hosted a 24-hour hackathon focused on the ARC AGI 2 challenge, a benchmark that tests a model's ability to solve complex reasoning puzzles from minimal demonstrations. A team of ML engineers from 2K Games and Google DeepMind participated in this event, aiming to build a system that could far surpass the performance of existing frontier models on this difficult task.

The winning solution, developed on Strong Compute's Burst Workstations and ISC cluster of H100 GPUs, combined One-Shot Reinforcement Learning with an AlphaEvolve-style search to create a self-improving system. This approach allowed the team to solve approximately 80% of the training puzzles, a success rate that dramatically exceeded the single-digit scores typical of other leading models. Strong Compute's infrastructure enabled rapid scaling and provided crucial support throughout the event.

2K Games