Learning: what the theory says, and what has been measured
You remember the work you did with the material, not the material itself
Too easy or too hard, the track stays shallow; in the middle it carves deepest
Everyone knows the feeling. You read the chapter twice and nod along the whole way, and a week later almost nothing is left. But the problem you sat and struggled with, you still remember. The idea I work from is that this is no accident. You do not keep what you are shown. You keep the work you did yourself when you had to choose between several possible answers. On this page I explain the idea, show what I have measured in language models, and say just as plainly what I have not been able to show.
A track is laid when a choice is made
In friction theory, a system that learns always has several possible answers in play at once. They compete, and in the end the system picks one of them. Holding several options open and then closing them again costs something. The idea is that this very work leaves a track. Next time the track is already there, and the answer is easier to find.
A system that carries traces of its own history is called hysteretic. The word is borrowed from physics, and I use it as a useful simplification, not as a description of what nerve cells do. The point is simple: a system that never commits to an answer lays down no track. A system that chooses against alternatives pays for it, and the payment is what gets remembered.
You remember the work, not the material
This is the idea Paper 16 calls encoding-through-loading: what sticks is the processing the system did itself, not the information it was given. Input that does not open a race between possible answers leaves no trace.
The idea has a well-known ancestor. Craik and Lockhart showed in 1972 that deeper processing gives better memory. The difference lies in what counts. Here it is not depth as such, but whether the processing opened a race that had to be settled.
It also explains why being tested works better than rereading. When you have to retrieve the answer yourself, you have to choose between the possibilities that come to mind. That choice is the work, and the work is the track. Rereading opens almost no competition.
Three ways teaching goes wrong
If learning is the work of settling a race, teaching can go wrong in three ways, and most people who have taught will recognise all three.
- Too much at once. The material is bigger than the system can sort through in time.
- Too thin. The right information is there, but so diluted that the race never gets going. Then there is no work to remember.
- Never a decision. Several readings are held open all the way through, with nothing ever closing them.
Each has its own fix. Cut down, concentrate the material, or make some of the choices for the learner. And one more thing follows: do not explain everything. When everything is explained, the work has been taken away from the person who was meant to learn.
What the language models show
You cannot look inside a brain at the moment a choice is made. In a language model you can watch the competition between possible next words, word by word. That is why I use language models as a mechanical mirror. The argument does not rest on them. It rests on half a century of research on people. But the models show the mechanism at a resolution you cannot otherwise get. Two measurements matter most here.
Training facts in is not the same as having them in front of you
In Paper 2B a language model was given 47 invented facts in two ways. Once they sat in the prompt, right in front of it. The other time they were trained into the model's weights. With the facts in front of it, the model gave the right answer 88 per cent of the time. Trained in over 15 passes, it reached 46 per cent, and even after 100 passes it was still 16 to 28 percentage points behind.
The most interesting part is what the training did along the way. The competition between possible answers was squeezed together until the model chose its answer with near-total certainty. The uncertainty that could otherwise be seen in the competition was pressed so far down that it no longer showed. When the same facts were trained in several different phrasings, with the same amount of computing, the competition was squeezed less. The variation protected it.
This is measured in one model, and it says nothing about how large a similar effect would be in a person. But the pattern is easy to recognise: pouring the material in is not the same as working with it.
When help harms the one who already knows how
Kalyuga, Ayres, Chandler and Sweller described in 2003 what is called expertise reversal: support that helps beginners can harm those who already know how. A recent meta-analysis (Tetzlaff and colleagues, 2025) puts the effect at about +0.5 for novices and −0.4 for experts.
In Paper 4B, three language models of different sizes were given the same task with 0, 1, 3 and more worked examples in front of them. The smallest (Qwen2-1.5B) stayed flat at 4 to 6 per cent. It could not solve the task, with or without examples. The middle one (Qwen2.5-7B) rose from 31 to 43 per cent with one example and then held steady around 41 to 46. The largest (Llama-3.3-70B) fell: 73 per cent correct with no examples, 50 with one, 61 with three and then around 62 to 64. The model that best knew how to solve the task was the one a single example disrupted most.
The paper's explanation is that the example reopens a race between strategies that the model had already closed. The competition between possible words was measured in the same runs, and it peaked exactly where the number of correct answers dropped.
Three caveats belong here. The pattern did not hold on a different kind of task, so it applies to the task it was measured on. The numbers move a few percentage points from run to run. And whether the same mechanism is at work in people is not settled. It is a prediction that has to be tested in people.
Bjork, spacing, and what I have not been able to show
Robert Bjork and his colleagues have shown that what feels difficult while learning often gives better memory afterwards. Retrieving instead of rereading. Spreading practice out instead of cramming it. Mixing topics instead of taking one at a time. This is solid research on people, and it stands on its own.
Paper 16 proposes one shared mechanism behind these difficulties: each of them forces a choice against an extra kind of alternative. Spaced practice becomes a race against forgetting, where the track has to be found again while it is fading. Mixed topics become a race against confusion, which sharpens the ability to tell things apart. That is a proposed mechanism, not a new finding about people.
I have also tried to find spacing in language models, and it did not work. In Paper 6 the spaced-training experiments came back null. The models sat at the floor at every spacing, so the test could neither confirm nor rule out anything. Paper 6 records it as a prediction that failed in the form the theory had at the time, not as a finding for or against Bjork. The question of spacing in language models is therefore undecided. There is also a reason for caution: a language model does not remember from one conversation to the next, so phenomena that are about the time between two practice sessions are not something it can simply mirror.
What would knock it down
An idea is only worth something if it can be wrong, and Paper 16 says what would knock it down. If plain repetition of the same phrasing sticks as well as varied repetition, measured after one week and after four weeks in large pre-registered studies, the prediction falls. If the window where the resistance is just right turns out to be symmetric, so that overshooting costs the same as undershooting, another one falls. If both fail, it is not a detail that needs correcting but the account as a whole that is in difficulty.
What it means if it holds
For you who teach or build courses: the target is the window where the resistance is just right. Too easy, and the race never gets going. Too hard, and it breaks down. It is in the middle that learning sticks. Where the middle lies moves with the material and with the learner, so the theory gives you a language to think in, not a formula.
For compliance and workplaces: the same mechanics explain why information-heavy compliance courses rarely change behaviour. More text does not build the route. I unpack this on Compliance is behaviour, not information.
For Dunning-Kruger: what happens to confidence while competence is still arriving, measured in language models, is on the Dunning-Kruger page.
The full account is in The Physics of Learning (Paper 16), and the expertise-reversal measurements are in You remember the work, not the material (Paper 4B). If you want the picture underneath all of this, What is a race? starts with water: tanks, pipes, and the channels in the sand that learning leaves behind.