A mathematical theory of balancing relational generalization and memorization
The paper studies how learning systems can follow a rule while still remembering exceptions, and why that balance is hard.
Intelligence analysis by GPT-5.4 Mini
This arXiv paper introduces a new task for testing whether models can learn a relation and also store an exception. It argues that the balance between generalizing and memorizing depends on representation geometry, and shows the pattern in finetuned language models.
A learner can be good at following a rule, like “A comes before B, and B comes before C.” But sometimes there is one special case that breaks the pattern.
This paper asks how a machine can remember both things at once: the big rule and the oddball exception. That is a bit like a student learning the highway rules, but also remembering one road that has a strange detour.
The researchers say the answer depends on how the information is arranged inside the machine’s mind. When they tested language models, the models often got the general rule right but still made the kinds of mistakes the theory expected.
Analysis
What the paper studies
The authors focus on a core learning problem: how a system can learn a general rule without losing track of the exception. They argue that this tradeoff has been hard to study because there have not been enough task setups designed specifically for it.
To fill that gap, they introduce a new benchmark called transitive inference with exceptions. In plain terms, it checks whether a learner can infer ordered relations while also remembering that one item breaks the rule. That makes the task stricter than ordinary transitive inference, where the system only needs to follow the chain and does not need to store a special case.
What they find
They analyze a simple but mathematically tractable model, kernel ridge regression, across a wide range of representation shapes and task settings. Their theory shows that the model can, in fact, balance relational generalization and memorization. But success is not automatic. Unlike the easier version of transitive inference, this harder task depends strongly on the geometry of the representations.
The paper frames that as a mechanistic explanation for why the task is more difficult: the way examples are arranged in representation space changes whether the model can keep the rule and the exception separate enough to learn both.
Why it matters
The authors then test the ideas in pretrained language models that are finetuned on ordered relations. Those models do learn the transitive rule, but they also make systematic mistakes that match the theory’s predictions. The paper’s broader point is that AI systems can look like they understand a rule while still mishandling exceptions, and that new task designs are needed to probe this reliably.
Key points
- The paper studies how models balance rule-based generalization with memorizing exceptions.
- It introduces a new task, transitive inference with exceptions, to test that tradeoff.
- The authors analyze kernel ridge regression across many representation geometries and task settings.
- Their theory says the harder task is sensitive to how representations are arranged.
- Finetuned language models show systematic mistakes that match the theory’s predictions.
