The Astra model demonstrates the ability to perform inference similar to belief propagation without using a chain of reasoning. This was discovered during testing of the model on logical reasoning tasks.
Testing showed that Astra uses token fillers for iterative convergence to exact marginal probabilities. This was achieved in 1 hour and brought 22 points.
Such results suggest that Astra can perform complex logical operations without explicit use of a chain of reasoning.
This discovery may have significant implications for the development of artificial intelligence models and their application in various fields.