Let’s start with the fact that the entire system was written from scratch so don’t hit me too hard with your keyboards when I open source it :D

My system is essentially split into several semi-independent modules communicating with each other when out-of-domain actions are needed. One of these modules is what I call the “logic reasoning module” and it is essentially a bunch of narrow specialized agents serving as a glue between the task and the bag of algos I found in the wild. One of its purposes is to apply formal logic to check whether the text given to it is correct, and to formally infer properties of some parts of the text (for example, if the text mentiones a certain door, and the system needs to ensure that the door, given its previosuly learned properties and a textual description, is indeed a door and has no undesirable properties such as being a broken door or a hard-to-open door). Another thing this module does is decision making. Agent generator creates state evaluation agents and all the other necessary entities from blueprints and then sends their actor references to the algorithm, such as mcts for example.

But I gave up on making this module work as I wanted and came up with a reasoning habit module instead. That one is a meta-module that essentially keeps track of the entire set of system activities and tries to detect any sorts of patterns, and its sub-module then tries to create a “shortcut”. The thing is, these learned patterns have individual scores w.r.t. the skill they were made for. These patterns essentially compete with each other for the right to be used in their respective cases. Basically, a schizo form of a reinforcement learning approach.

There’s much more to it than what I already described but I’m too sleepy so nope. And yes, you don’t really need large LMs for it to function, like at all. Yes, they will give you somewhat better result, but their cost is a big oof.

P.S.: I use knowledge graph with a few extensions (like that one that resembles frames), and this graph also has temporal component and a simple node level version control. I just ran out of hobbies and I really wanted to see how exactly would my attempt to build that all fail so here I am :D