It works by training a weight matrix (N-to-N) on different solutions. For example, in sudoku you're training a weight matrix for every space to every other space (81 c 2), with 9x9 numbers (~240K weights). So then the "energy" for every number in a space is W_ij(a, b) where ij is a pair of spaces, and (a, b) is a pair numbers you want to swap (e.g. 4->2 in the space). Pick the lowest energy move each time (argmin of this whole space), and iterate.
This has a few limitations. First, its discrete and has a limited space definition, getting this to work for arbitrary reasoning seems hard (what size/shape graph would you use for arbitrary reasoning). Second, the ruleset is defined apriori (e.g. you need to know you have a 9x9 grid and 9 numbers, the rules of the game are implicit in the checker). Third, the actual search algorithm is somewhat bespoke to task (less of a factor, maybe use generic graph coloring as the basis of any problem), but for something like agentic reasoning its not clear how this would be adapted. For example with mazes, we need to know what constitutes a valid path and a solved maze, but for language/code rules like that are hard to formulate.
Fourth, this doesn't admit of an easy way to use "fuzzy reasoning" that would be required for integration with any kind of human language or image-based system, where continuous vectors are useful (the graph edges and colors are on/off or discrete).
There are ways to adapt, but this approach fits pretty solidly in the "deterministic, fixed problem size" realm.
In addition, for training you need correct solutions and counter-examples (strongly labeled). For something like "fuzzy logic" training or "natural language" coming up with clear examples and counter-examples would be challenging to fit into this model. LLMs are only possible by leveraging vast quantities of unlabeled data. Not saying impossible, but challenging. For example, if training on human code, what constitutes a counter-example? You basically already have to have a method of getting correct and sufficiently incorrect code just to train this type of system, and its unclear how to represent those changes.
For all of these problems, given the constraints and amount of bespoke logic required, solving with dynamic programming is far faster (including arbitrary graph coloring problems). I only say this to mean that if we could solve all the problems I mentioned above and get arbitrary reasoning into the form of graph coloring, it would still probably be more efficient just to solve that with good old dynamic programming / recursion than with a model like this that needs O(N^2) params versus the graph and needs millions of training examples for different domains. They brag about solving sudoku, but a CPU-based approach to solving it the good old fashioned way on is probably 10-100x faster. If we were using an H200 like they did, we could easily solve the whole sudoku dataset in less than a second.
But it _is_ a very interesting approach, partly because the energy strategy encodes good / likely guesses, so that is kind of interesting vs dynamic programming which is kind of a brute-force approach, so maybe for very large spaces it would be more efficient (but then you have the issue of the parameter space blowing up).