Seed IQ Achieves 100% Score on ARC-AGI 3

Seed IQ achieved a 100% score on the ARC-AGI 3 benchmark by leveraging Active Inference rather than LLM scaling or brute force methods. On some levels, it recorded superhuman performance with scores of 115%, 2 to 3 times the human baseline.

This system works by inferring Priors such as environmental invariances, constraints, symmetries, and object relationships to collapse the Search space. Through this, instead of simply memorizing or sampling data, the model identifies the dominant structure of the task and moves directly to the solution manifold.

The key technical features are as follows:

• No use of Zero tokens or Reinforcement Learning (RL)
• Combination of Topological perception and physics-based dynamics
• Deterministic and fast performance implementation

It is not currently registered on the official leaderboard, which was stated to be because they refused to disclose the codebase and commercialization rights in order to protect proprietary IP.