Learning the Past Tense in AI

Arshavir Blackwell, PhD

This is Inside the Black Box and I'm Arshavir Blackwell.

Arshavir Blackwell, PhD

There is a moment every English speaking toddler goes through that has launched a thousand academic papers. The child, who has been saying went correctly for months, suddenly starts saying goed. And taked. And bringed. Then, months later, they quietly go back to went.

Arshavir Blackwell, PhD

Plot a child's accuracy on irregular verbs over time and you get a famous dip, a U shape. Right, then wrong, then right again. It looks like the child learned a rule ("add e d to make the past tense"), got so enamored of it that they over applied it to exceptions they used to know, and only later learned to hold the rule back.

Arshavir Blackwell, PhD

That little dip is one of the most argued about facts in the science of language. And it is the setup for a question I have been chasing: do AI language models do the same thing, and if not, what are they doing instead?

Arshavir Blackwell, PhD

The fight. Two camps have been swinging at each other over the past tense since the 1980s.

Arshavir Blackwell, PhD

On one side, the rule people (Pinker and Prince, 1988, is the canonical shot). Their claim: the mind has two systems. A symbolic rule, literally "add e d", and a separate memory for the exceptions like went and ate. Two routes. Overregularization happens when the rule fires before memory retrieves the exception. Kids say goed because the rule wins the race.

Arshavir Blackwell, PhD

On the other side are the network people (Rumelhart and McClelland, 1986). Their claim: you do not need a rule at all. A single associative network, trained on examples, will reproduce all of this, including the U shaped dip, with no explicit rule anywhere inside it. Their headline result was exactly that: a neural network that produced the child's goed stage on its own.

Arshavir Blackwell, PhD

The debate never really resolved, and here is the frustrating reason: behavior underdetermines mechanism. Both theories predict roughly the same outputs. You can watch a child (or a network) say goed until you are blue in the face and never see which story is true inside the head doing the talking.

Arshavir Blackwell, PhD

Why now. Large language models change the game, not because they settle the debate by fiat, but because of one boring, decisive fact: we can open them up.

Arshavir Blackwell, PhD

An LLM is a distributed network that plainly learns inflection. But unlike a human brain (or a 1986 model on a lab tape drive), a modern open model comes with its internal activations, its weights, and, crucially, snapshots of itself taken all the way through training. We can watch it learn frame by frame. We can read its internal state. We can reach in and perturb it. We can switch pieces off.

Arshavir Blackwell, PhD

So I ran the past tense question in silico, and I ran it with five different methods, because if there is one lesson interpretability keeps teaching, it is that any single probe can fool you.

Arshavir Blackwell, PhD

This first post is about the simplest method, and the one that speaks most directly to that childhood U: watch the model learn, and see if it dips.

Arshavir Blackwell, PhD

Watching a model learn. I used Pythia 410M, an open model that helpfully publishes 140 odd checkpoints from across its training run. At each checkpoint I measured two things, using a length fair scoring method (so a two token answer is not unfairly penalized against a one token one):

Arshavir Blackwell, PhD

First, Overregularization: how often it prefers goed over went. Second, Competence: how often it gets ordinary regular verbs right (walk to walked).

Arshavir Blackwell, PhD

Here is what came out: Look at where the goed peaks. It is at the very beginning, when the model is worst at everything, 80 percent goed at 38 percent competence. From there it just fades away, smoothly, as the model gets good.

Arshavir Blackwell, PhD

There is no dip and recovery. There is no "get good, then start saying goed, then recover." The model does it backwards from a child.

Arshavir Blackwell, PhD

Overregularization and regular verb competence across training, for two model families. In a child, overregularization would rise after competence climbs. In both models it does the opposite: overregularization is highest when competence is lowest, and only falls. No U shape.

Arshavir Blackwell, PhD

Backwards from a child. A child says went correctly first (rote memory), then over applies the rule, then recovers, so the goed stage comes after competence.

Arshavir Blackwell, PhD

The AI is the mirror image. It starts out slapping e d on everything, that is its lazy default, because e d is the most common ending it sees, and it learns the exceptions on top of that, gradually, as it encounters them more. Goed in the AI is not a rule being over applied by a competent learner. It is a beginner's shrug that gets slowly corrected.

Arshavir Blackwell, PhD

This already matters for the old fight. Remember, Rumelhart and McClelland's whole selling point was that their network reproduced the child's U. Modern networks do not, which, ironically, lines up with an old Pinker and Prince objection that the 1986 U was an artifact of how the network was fed its data.

Arshavir Blackwell, PhD

Does it hold up? One model is an anecdote. So I repeated it on OLMo 1B, a model from a different lab (AllenAI), trained on different data, that also publishes checkpoints. Same story: goed is highest at the least competent snapshot and drops to zero once the model is fully competent. No hump after mastery.

Arshavir Blackwell, PhD

It has a limitation: OLMo does not save snapshots as early as Pythia does, its first public checkpoint is already fairly competent. So OLMo cannot show the very first e d on everything spike; it can only confirm the key negative, that overregularization does not rise after competence, the way it does in children. Two families, same conclusion, one with a fuller view than the other.

Arshavir Blackwell, PhD

Where this leaves us. So the first finding is a clean no: the models do not learn past tense the way children do. They acquire the rule shaped default first and memorize the exceptions over it, the reverse order, and they never produce the human U.

Arshavir Blackwell, PhD

But notice what that does not tell us. "It does not learn like a kid" is not the same as "it has no rule." Maybe the AI still has a genuine, productive e d rule; it just acquires it in a different order.

Arshavir Blackwell, PhD

That is the question for Part 2: is there actually a rule in there, a clean, general add e d operation, or is the model mostly leaning on memorized word forms, with only a thin, leaky kind of generalization? The answer, it turns out, is neither of the tidy options. There is a beautiful 1958 experiment, involving a made up creature called a wug, that cuts straight to the heart of it. And then we will go looking inside the machine for the rule itself.

Arshavir Blackwell, PhD

It does not go the way the textbook predicts. Next: Part 2, Rule or Rote? The wug test, and the hunt for a past tense machine inside the model.