What limits reasoning capability beyond math and code?

Reasoning models are trained largely via RL on tasks where the reward can be rule-verified — which is why existing reasoning datasets cluster in narrow domains with short, easily checked solutions (math, coding). But most reasoning across broader domains is open-ended. NaturalReasoning's argument is that the binding constraint for scaling reasoning beyond math and code is not a better training method but the supply of diverse, challenging questions.

It addresses that with a scalable generation approach producing 2.8M questions (with reference answers) curated from pretraining corpora, spanning physics, computer science, economics, social sciences, and more — selected for diversity and difficulty relative to existing datasets. The evidence that the data, not the method, is the lever: distillation experiments show consistent improvement on reasoning benchmarks as data size scales, and the dataset also enables unsupervised self-training via external reward models or self-rewarding.

The methodological keeper is that question difficulty and breadth are first-class inputs to reasoning capability — "reasoning in the wild" requires more deliberate thinking than narrow verifiable datasets elicit, and capability transfers from a strong teacher through these questions. This complements Can models improve themselves on tasks without verifiable answers?: that note shows a small demonstration set can unlock general reasoning; NaturalReasoning shows a large question set scales it. Together they bracket the question-supply problem from the demonstration and scale directions.