[KongchangAI]
ModelClassical Energies Reasoner with Adaptive Minimization

CRAM

CRAM is an inference architecture based on energy models that replaces the iterative solver of recurrent models with a learned parameterized energy function, enabling fast training and inference. It features early stopping and adaptivity of computational effort.

Timeline (last 90 days)

Nov 13

On Sudoku and maze tasks, CRAM trains at least 1000 times faster than recurrent models.

Unverified50%
Nov 13

On Sudoku and maze tasks, CRAM reduces inference FLOPs by at least 282 times compared with recurrent models.

Unverified50%
Nov 13

After training for 42 milliseconds on a single GPU, CRAM solved 99.996% of Sudoku-Extreme test puzzles.

Unverified50%
Nov 13

The key innovation of CRAM is that it no longer hand-writes the energy function but instead parameterizes the energy and learns it in a data-driven way, so that the minima of the energy landscape correspond to the correct solutions of the puzzle; at inference time it only needs to iterate along the direction of energy descent, without backpropagating gradients through the solver.

Unverified50%
Nov 13

CRAM has a built-in early-stopping mechanism: iteration stops when the energy no longer decreases.

Unverified50%
Nov 13

CRAM's computation adapts to problem difficulty: easy instances fall into minima faster.

Unverified50%
Nov 13

The weights learned by CRAM are interpretable and correspond to the puzzle's constraints.

Unverified50%

All Facts (7)

Source Articles