Poetiq shatters ARC-AGI 2 benchmark at half the cost
5 days ago
- #AI
- #Machine Learning
- #ARC-AGI
- Poetiq's system has been officially verified to outperform existing methods on the ARC-AGI-2 Semi-Private Test Set, setting a new state-of-the-art.
- The system achieved a 54% success rate at $30.57 per problem, surpassing the previous best score of 45% at $77.16 per problem.
- Poetiq's meta-system optimizes solutions by leveraging existing frontier models without the need for building or fine-tuning new models.
- The meta-system learns from each task it solves, improving over time, and can be integrated into larger AI systems.
- Poetiq is exploring the potential to solve long horizon tasks by enhancing knowledge extraction mechanisms without model tuning.
- The team consists of 6 experienced researchers and engineers from Google DeepMind, focusing on AI reasoning and knowledge extraction challenges.