The problem
Imagine cards arranged in a hidden ranking. You learn that A outranks B and B outranks C, but never compare A with C. How do you work out that A outranks C, and revise that conclusion if the ranking changes? The question concerns a consistent order; winning two games in a row would not by itself guarantee this relationship.[1] [2] [3]
Imagine learning neighboring pairs in a hidden order: A above B, B above C, C above D, and D above E. You can be asked to compare B with D even though that pair was never taught. How does the mind construct that answer, and what changes if E moves to the top? This is transitive inference within a learned linear order.
One account retrieves and links remembered pairs. Another assigns each item a learned value related to its training outcomes. A third represents positions in a common ranking. Adult experiments on changed orders test the predictions of value updates; child studies compare generalization with pair memory. A computational neuron model shows how local learning could build ranks under its assumptions about how items are represented by neurons. These are candidate explanations, rather than evidence that everyone uses the same neural rule.[1] [2] [3]
Scope
Human inference of untrained pairwise comparisons in learned linear orders, including developmental differences and adaptation when the order changes.
What remains open
People can judge pairs they were not shown. They might link remembered comparisons, learn how often each item wins, or build a ranking. Child and adult tasks support different mixtures of these strategies, and a neuron model offers one possible implementation. The results do not establish which mechanism people use across tasks.[1] [2] [3]
Older children performed near the ceiling in one study, making it harder to separate competing explanations. Adult results with deterministic feedback leave open how people distinguish a genuinely changed order from unreliable feedback. The local rank-learning model produces predictions about human neural representations that still need experiments; its mathematical convergence does not establish those representations in people.[1] [2] [3]
Why it matters
A common ranking and a collection of learned pair outcomes can agree on familiar questions yet react differently to one changed relation. That difference matters for how knowledge is reused: a learner can extend a few examples to unseen comparisons, but may also spread an outdated assumption through connected judgments. Explaining when this happens would clarify how inference and correction work together, rather than treating each answer as an isolated memory.[2] [3]
Definitions
2 terms
Transitive inference
Using linked comparisons in a consistent order to judge a pair that was not directly compared; for example, A above B and B above C imply A above C.[1] [2] [3]
Learned value
An item’s learned score related to its training outcomes. Such scores can guide comparisons without being an explicit stored ranking.[2]
Sources
- Generalization from early to middle childhood
Notes on Generalization from early to middle childhood
Children’s ordered-comparison performance supports developmental differences; near-ceiling older performance limits discrimination between models.
- Asymmetric learning and adaptability to changes in relational structure during transitive inference
Notes on Asymmetric learning and adaptability to changes in relational structure during transitive inference
Adult experiments explain direction-dependent adaptation through biased value updates, while leaving noisy feedback and broader model comparisons open.
- The inherent capacity of neurons to learn order relations and support abstract reasoning
Notes on The inherent capacity of neurons to learn order relations and support abstract reasoning
A local plasticity model learns ranks and explains several effects; explicit human rank representations remain an experimental prediction.