Is a model's math skill one single ability, or many separate skills that just happen to move together?

This explores whether a model's math ability is one thing or several things wearing one number. The corpus has no study aimed squarely at this, but it has a useful tension. Some notes find that what looks like a split is unified underneath. Others find that what looks like one ability is really several pulling apart.

The case for 'unified' comes from a surprising place. In reinforcement learning on math problems, researchers have long assumed a trade-off: a model either explores new solution paths or exploits the ones it already knows. But when you look at the model's internal states instead of its output words, the two barely correlate, and both can be improved at once, which produced large gains on a Chinese university entrance math exam Is the exploration-exploitation trade-off actually fundamental?. The lesson goes beyond RL. Splits that appear at the level of output text can disappear when you measure what's happening inside the model.

The case for 'separate' is just as strong. A study using simple cipher puzzles broke chain-of-thought performance into three independent ingredients. One is how likely the answer is in general. One is memorization of patterns seen in training. The last is real step-by-step reasoning that picks up errors as it goes. Output probability alone moved accuracy from 26% to 70% What three separate factors drive chain-of-thought performance?. So a single 'math score' can blend memory, prior expectations and genuine reasoning in proportions you can't see. Agent evaluation shows a similar pattern: capability splits into axes that rank models differently, so any single score misleads Does a single benchmark score actually predict agent readiness?.

The most unsettling doorway shows that even perfect accuracy tells you little about how a skill is organized inside. Models can contain every feature a task needs while their internal structure stays fractured. That breakage only shows up when inputs shift Can models be smart without organized internal structure?. So 'unified vs. separate' may be a question about the inside of the model, and benchmarks can't answer it. Another note makes a related point: the strong math scores of a small 3B model come from how it was post-trained, not from its size, and those results hold only where answers can be checked Can small models match frontier reasoning without massive scale?. That suggests 'math capability' is partly something the training process builds, not a fixed property of the model.

The takeaway you might not have expected: the answer depends on what you measure. At the level of output words and benchmark scores, math ability looks like separable pieces. In the model's hidden states, some of those apparent splits merge, and seemingly solid skills can turn out fractured. If you meant something narrower, such as how difficulty or accuracy is spread across math benchmarks, the corpus doesn't cover that directly.