The Distinction in One Paragraph
Raymond Cattell proposed in the 1940s, and developed with John Horn through the 1960s, that general intelligence is better described as two broad abilities that behave differently.
Crystallised intelligence (Gc) is stored knowledge and learned procedure — vocabulary, general information, taught methods. Fluid intelligence (Gf) is the capacity to solve a problem with nothing stored to draw on, leaning on working memory and pattern detection.
Verbal reasoning falls cleanly on the crystallised side. Abstract reasoning falls cleanly on the fluid side. Numerical reasoning does neither, and that is the interesting part.
Why Numerical Reasoning Splits Down the Middle
Take apart a single data-interpretation item and the two components separate visibly.
The crystallised half
- The operations themselves. Knowing how to compute a percentage change is taught knowledge, retrieved rather than derived.
- Chart and table conventions. Recognising what a stacked bar or an indexed series is showing is familiarity, not reasoning.
- Number sense. Knowing immediately that 17% of 240 is near 40 is a stored feel for magnitudes, built by exposure.
The fluid half
- Holding intermediate results. Carrying one figure in mind while scanning the table for the next is working-memory load, and it is the component that collapses first under fatigue.
- Deciding what to compute. Choosing the operation from an ambiguous question is a novel-problem step, not a retrieval one.
- Sequences. Number-series items are structurally closer to abstract reasoning than to arithmetic — the task is finding an unstated rule.
The Evidence From the Test Batteries Themselves
The strongest evidence that numerical reasoning is a hybrid is that professional test builders could not agree on where to file it.
Wechsler's Arithmetic subtest sat inside the Verbal half from 1939 through the third edition — verbal in delivery, treated as verbal in scoring. The fourth edition in 2008 moved it to the Working Memory Index.
Nothing about the subtest changed. What changed was the recognition that its behaviour matched working memory better than it matched stored verbal knowledge. A task can sit defensibly under two constructs for fifty years only if it genuinely straddles them.
Why It Is More Fluid Than Verbal Reasoning
The comparison that makes this concrete is against a vocabulary item.
A vocabulary item is pure retrieval. You either hold the word or you do not, and the working-memory demand of answering is near zero.
A data-interpretation item cannot be answered by retrieval at all. Every figure in the table is new, and the reasoning has to be constructed on the spot from material you have never seen — which is the definition of a fluid task, even though the tools being applied are learned.
What the load actually is
The bottleneck in numerical work is rarely the calculation. It is that a two-step item requires holding a result while performing a lookup, and the lookup is exactly what displaces the held value.
This is why candidates report a specific failure mode — computing correctly, then realising they can no longer remember which figure they were computing with.
The Lifespan Pattern Sits in Between
If numerical reasoning were purely crystallised, it would behave like vocabulary and keep improving into middle age. If it were purely fluid, it would peak early and decline like matrix reasoning.
The observed pattern is intermediate, and it is what a hybrid predicts. Longitudinal work — the Seattle Longitudinal Study being the most cited — consistently finds fluid measures peaking in early adulthood and declining thereafter, with crystallised measures holding up far longer.
Numerical ability holds up better than abstract reasoning and less well than vocabulary, which is exactly the position its component structure implies.
The practical version
An experienced analyst at fifty typically has better judgement about what a number means and slower raw manipulation than they did at twenty-five. Both halves of that sentence are the Gf/Gc split showing through.
Does the Fluid Component Mean It Cannot Be Improved?
No — but it does mean improvement has a shape, and knowing the shape saves wasted effort.
What responds well
- Arithmetic fluency. Making the operations automatic frees working memory for the reasoning. This is the single highest-return investment, and it is pure practice.
- Format familiarity. Recognising chart types and question phrasings removes a decoding step that was costing seconds per item.
- Estimation habits. Learning to sanity-check magnitude before calculating catches the extraction errors that are otherwise invisible.
What responds poorly
The working-memory capacity itself. The training literature has repeatedly found that people improve at the trained task while transfer to untrained reasoning measures is weak or absent.
The useful move is therefore not to enlarge the capacity but to stop wasting it. Every operation made automatic is capacity returned to the reasoning — which is why fluency drilling outperforms working on harder problems.
Which Plateaus First?
Numerical, and by a clear margin.
Because a large part of numerical performance is fluency with a small fixed set of operations, most of the available gain arrives in the first several hours of focused practice and then flattens. The syllabus is genuinely short, and once the operations are automatic there is little left to automate.
Verbal reasoning has no such ceiling, because vocabulary and syntactic familiarity keep accumulating for as long as you keep reading. Decades, in principle.
For anyone deciding where to spend preparation time, that asymmetry is the answer: numerical practice pays back fast and stops, verbal practice pays back slowly and continues.