Energy-Based Reasoners: 385x more efficient inference
A method using a learned energy to refine, stop, and select inference achieved over 99% accuracy on Sudoku-Extreme.
Key points
- •The model has 1.8M parameters and achieved over 99% accuracy on Sudoku-Extreme.
- •It reduces inference FLOPs by up to 345x compared to EqR at comparable accuracy.
- •A single learned energy handles the refinement, stopping, and selection of the inference process.