Training AI on examples where its answers vary most could be the secret to reasoning that actually scales.

This explores whether picking training examples by their reward variance — a curriculum that favors prompts where the model's outcomes still disagree — actually makes reasoning training scale better. The corpus has a fairly direct answer to the mechanism, and a more skeptical answer to the word 'reasoning.' On mechanism: reward variance is a real signal, not a bookkeeping detail. When within-prompt reward variance is low, the useful task gradient weakens and regularization takes over, so the policy drifts toward generic, input-agnostic templates — the same answer shape regardless of the question. Selecting high-variance prompts before each update (SNR-Aware Filtering) recovers performance across tasks and, notably, keeps working as you scale Why do language models collapse into generic templates?. So as a *curriculum*, reward variance targets exactly the prompts where the model has something left to learn and stops wasting updates on ones it has already saturated.

But here's the twist that makes this more interesting than a training trick: what does that saturation actually buy you? A cluster of work argues RL on verifiable rewards mostly *activates* strategies the base model already has rather than teaching genuinely new reasoning — a single training example can suffice to switch behavior on, and even spurious rewards work nearly as well as correct ones for a well-pretrained model What does reward learning actually do to model reasoning?. The same theme runs through the finding that base models already carry latent reasoning that minimal training merely elicits — post-training *selects* reasoning, it doesn't create it Do base models already contain hidden reasoning ability?. Read together, a variance curriculum is best understood as an *elicitation efficiency* tool: it finds the prompts where the model's latent competence hasn't yet been surfaced and concentrates learning there. That reframes 'scalability' — you're scaling how efficiently you unlock existing capability, not necessarily pushing the capability ceiling upward.

Which raises the honest ceiling question. Variance is a scalar; once a prompt's outcomes agree, the signal goes silent even if the model is agreeing on a *wrong* or shallow answer. That's precisely the gap other work flags: numerical rewards lack information about *why* a solution failed, and models stuck on plateaus break through when given chain-of-thought critiques instead of just a number Can natural language feedback overcome numerical reward plateaus?. So a pure reward-variance curriculum inherits the blind spot of scalar rewards — it can tell you *which* prompts are still contested, but not *how* to improve on them. The richer-signal alternatives (generative judges that reason about reasoning steps, reward models that themselves deploy test-time reasoning before scoring) suggest the frontier is making the reward signal *carry more structure*, not just filtering on the variance of a thin one Can judges that reason about reasoning outperform classifier rewards? Can reward models benefit from reasoning before scoring?.

So the layered answer: yes, curriculum-by-reward-variance improves *training* scalability — it prevents template collapse and keeps gradients informative as you grow the run — and that's a genuine, measured win. But it scales the *efficiency of eliciting* what's already latent more than it scales raw reasoning capability, and its scalar nature caps how far it can carry you compared to feedback that explains failures rather than merely detecting disagreement. The most interesting thing you might not have expected: the very same low-variance signal that a curriculum uses to *skip* a prompt is also the thing that, left unmanaged, actively *degrades* a model into generic templates — so variance filtering is doing double duty as both a curriculum and a collapse-prevention safeguard.