The ARC-AGI-3 competition on Kaggle challenges participants to develop agents capable of solving 8-bit style games. Participants are provided with a machine featuring a powerful GPU, 96 GB of VRAM, and 180 GB of RAM. The competition is structured in three stages: two milestones and a final round, with a one-month interval between each.
The author achieved 6th place on the public leaderboard. During the competition, participants were granted 30 hours of free access to high-performance hardware. The Qwen3.8 Next Flash model, which utilizes 512 experts, became the standard choice for most participants toward the end of the contest.
Initial attempts to improve performance through agent logic modifications, such as color scheme adjustments or parallel agent dueling, yielded no significant results. The author subsequently focused on inference optimization, specifically targeting server performance and vLLM integration to increase the number of parallel game lines.
By implementing techniques like KV int 4 memory quantization, reducing the expert count to 400, and utilizing stochastic drafts, the author increased processing speed to approximately 3400 tokens per second, enabling 110 parallel game lines. Total costs for the project included $30 for Colab and Claude subscriptions.
Despite achieving higher speeds, the author noted a trade-off between optimization and model accuracy. Technical challenges included Kaggle platform instability and server errors, which limited submission opportunities. Final efforts focused on overcoming performance plateaus through methods like agent swarms and multi-path decision-making.
