Models don't 'grow into' analogical reasoning with scale — they borrow it from patterns in their training data.

This reads the question as asking whether analogical reasoning is a stable skill models 'grow into' with more data and parameters, or a fragile artifact of training coverage. The corpus leans hard toward the second answer — analogical and reasoning-like behavior is far more sensitive to *what's in the training distribution* than to raw scale. Mechanistically, transformers do have a real analogy circuit: they align relational structure geometrically in embedding space and then apply a learned functor to map one domain onto another How do transformers perform analogical reasoning across domains?. But that circuit is assembled from patterns seen in training, not conjured from size.

The sharpest evidence that data proximity dominates comes from work showing reasoning failures track *instance novelty*, not task complexity. Models don't hit a wall at some difficulty threshold — they succeed on any reasoning chain that resembles training instances and fail on unfamiliar ones, because they've fit instance-level patterns rather than a general algorithm Do language models fail at reasoning due to complexity or novelty?. Chain-of-thought reinforces the point: it degrades predictably once you shift task, length, or format away from the training distribution, producing fluent-but-invalid reasoning Does chain-of-thought reasoning actually generalize beyond training data?. That's the signature of imitation — reproducing familiar reasoning *schemata* — rather than genuine abstract inference Does chain-of-thought reasoning reveal genuine inference or pattern matching? Why does chain-of-thought reasoning fail in predictable ways?. Even the length of a reasoning trace turns out to encode 'how close is this to my training data' rather than 'how hard is this problem' Does longer reasoning actually mean harder problems?.

Here's the twist that reframes the scale question: scale may not be the lever people think it is. Base models already carry latent reasoning capability, and five independent methods — RL steering, critique fine-tuning, decoding tweaks, SAE feature steering, RLVR — all just *elicit* what's already there. The bottleneck is elicitation, not capability acquisition Do base models already contain hidden reasoning ability?. A 3B model with the right curriculum and multi-domain RL matches much larger systems on verifiable tasks, so reasoning competence is a property of the post-training pipeline, not parameter count Can small models match frontier reasoning without massive scale?.

So the honest answer is that 'emergence' is the wrong frame. Analogical reasoning doesn't switch on at some scale — it's latent structure that becomes usable when training data and post-training coverage align relational patterns the model can reuse. What you'd notice you didn't know to ask: the same data can produce very different reasoning behavior depending on the verifier, optimizer, scaffold, and budget wrapped around it — the reusable unit isn't a dataset but the whole feedback interface What is the actual reusable unit of reasoning data?. And you can push reasoning *earlier* by treating chain-of-thought as an exploratory action rewarded during pretraining itself, which suggests where in the pipeline analogical structure gets planted is a design choice, not a fixed function of scale Can chain-of-thought reasoning be learned during pretraining itself?.