30% on AR-AGI-1 at $0.0007 per task (arxiv.org)

A groundbreaking AI model named BDH-CQ has been introduced, enhancing in-context learning through recurrent latent reasoning. This innovative approach allows the model to continuously update its memory as it processes inputs, solving queries using iterative computations within a high-dimensional latent space, all without verbalizing intermediate steps. The model was evaluated on the ARC-AGI-1 benchmark, achieving a remarkable 29.5% pass rate at a cost of just $0.0007 per task, marking a significant leap in cost-efficiency for AI reasoning models. The significance of BDH-CQ lies in its ability to break through the previously established accuracy-cost Pareto frontier, setting a new standard for how effectively AI can utilize resources while improving performance. With a 150 million parameter configuration, this model not only optimizes task accuracy but also enhances the learning process by identifying and applying transformations from demonstrations. This advancement showcases the potential for more efficient AI systems in practical applications, further pushing the boundaries of what is possible in the AI/ML community.