AI BriefingBDH-CQ Improves ARC-AGI-1 Performance
Key point
Introducing BDH-CQ, a system that demonstrates high efficiency on the ARC-AGI-1 benchmark through recurrent latent reasoning.
Details
BDH-CQ is a new reasoning system that utilizes Recurrent Latent Reasoning.
This system updates recurrent memory through demonstrations of unknown tasks and solves queries via iterative computation within a high-dimensional latent workspace. Intermediate reasoning states are not decoded into language, and memory, adaptation, and reasoning operate integrated within a single computational structure.
Key features and achievements include:
- Achieved 29.5% pass@2 on ARC-AGI-1 with a 150M parameter configuration
- Broke the existing cost-accuracy Pareto frontier with a cost of $0.00070 per task
- Capable of continuous memory updates and adaptation without parameter updates during inference
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.