Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model
One learned energy refines, stops and selects. Over 99% Sudoku-Extreme accuracy with 1.8M parameters, up to 345x fewer inference FLOPs than EqR at comparable accuracy.
One learned energy refines, stops and selects. Over 99% Sudoku-Extreme accuracy with 1.8M parameters, up to 345x fewer inference FLOPs than EqR at comparable accuracy.