Why cognitive architectures matter again in the age of LLMs

Large language models changed what machines can do with language. Give one a case file and it will summarise it, draft a reply and suggest a next step. It does this in seconds, and often well. It is fair to be impressed.

Now ask for something harder. Ask it to apply your policy the way your most careful senior colleague would, and to show how it weighed the evidence. You get fluent text. Fluent is not the same as faithful. A model’s account of its own reasoning is another piece of generated text. It may describe the process that produced the answer. It may not.

For casual use, that is fine. For decisions that need reasons, it is not.

An old idea

Long before language models, cognitive scientists and AI researchers asked a different question: what is the structure of a mind that solves problems? Their answers were called cognitive architectures. ACT-R and Soar are the best-known examples.

The shared idea is that intelligence comes from structure as well as from learning. There is a memory that holds what is known. There are goals. There are explicit rules for choosing what to do next. Because the parts are explicit, a model can be inspected, and it can be tested against how people actually behave.

The approach had limits. Knowledge had to be written by hand. Models were narrow and brittle. When learning from data began to scale, learning won, and cognitive architectures moved to the edge of the field. At the time, that was a reasonable verdict.

What has changed

Language models supply what the old systems lacked: broad knowledge, flexible language and tolerance for messy input. What they lack is what the old systems had.

Breadth lets a system read a document it has never seen. Structure lets it say what it did with that document, and lets a reviewer hold it to account. In regulated work, structure is the point. Three things matter most.

1. A place for rules that must hold. A policy is a constraint, not a suggestion. It belongs in a form the system cannot talk its way around and an auditor can read.
2. A model of the human task. What does the expert look at first? What does she recall from similar cases? Which strategy does she fall back on when time is short? A system that represents this can present information in the order she needs it.
3. A trace that is part of the system. The steps are recorded as the decision is made. They are not reconstructed as a story afterwards.

A model of the human task also works in the other direction. If the system represents what the expert is trying to do, her goals, her constraints and what she will not trade off, she can state that intent once, in her own terms, and the system can hold it from the first step to the last. She does not have to coax a prompt into saying what she means. That is the people-to-AI half of the link, and it rests on the same structure that makes the system explainable.

Two systems, one analyst

A compliance analyst at a bank reviews payments that have been flagged. Consider two systems that could help her.

The first is a language model alone. It reads the documents and writes a persuasive verdict. Nothing stops the same payment, described a little differently, from receiving a different verdict next week. She cannot tell why, and neither can anyone else.

The second combines parts. A language model reads the documents and extracts the facts. Explicit rules decide which checks apply. A model of how the analyst works decides what she sees first. The verdict arrives with its trace. If she disagrees, she can point to the step she disagrees with, and that step can be corrected.

On first sight the second system is no more impressive. Both are accurate on a good day. The difference shows on the day she has to defend a decision to an auditor, and only one of them can show its work.

What we do not claim

Cognitive architectures are not a cure. The old criticisms still stand. Hand-built knowledge is costly. A tidy structure can hide a wrong assumption. A model of how experts reason may simply be wrong about a particular expert. Each of these is an open question, and the answer has to come from testing with the people the model describes, not from admiring the diagram. Our research agenda lists the questions we are working on.

This is not a return to the past. It is an old idea with a new partner. Learning gives breadth. Structure gives accountability. Decisions that need reasons need both, and that is where NeuroLink AI works.