An independent researcher has achieved a new state-of-the-art score of 29.4% on the ARC-AGI v2 benchmark, surpassing the previous best of 25%, by replacing Python code generation with natural language instructions evolved via an evolutionary test-time compute architecture.
The system uses Grok-4 to generate plain English instructions for solving abstract pattern recognition tasks, which are then tested and refined through individual and pooled revision cycles. This approach yields a score of 79.6% on ARC-AGI v1 at $8.42 per task, making it 25 times more efficient than OpenAI's o3 preview.
The author argues that LLMs struggle with generalization due to "dead reasoning zones" where logic fails outside training distributions, and proposes evolving natural language instructions as a method to bridge the gap between human-like abstract reasoning and current AI capabilities.