When an AI model seems to learn a clever new way to reason, did training create it, or was it already hiding inside?

This explores how to tell whether a reasoning strategy was created by training or was already inside the base model waiting to be switched on. The corpus mostly points one way: much of what looks like new reasoning is latent. Five different methods (RL steering, critique fine-tuning, decoding changes, sparse-autoencoder feature steering and verifiable-reward RL) all pull reasoning out of a base model's existing activations. The conclusion is that post-training selects reasoning more than it creates it, and the bottleneck is elicitation rather than acquisition Do base models already contain hidden reasoning ability?.

The best tell for a latent strategy is that you can trigger it without teaching the model anything. Steering one identified feature matches chain-of-thought performance, and it does so by activating early in generation and overriding the surface instructions Can we trigger reasoning without explicit chain-of-thought prompts?. Four sandboxed 'cognitive tool' calls lifted GPT-4.1 on AIME2024 from 26.7% to 43.3% with no RL at all Can modular cognitive tools unlock reasoning without training?. Even behaviors like verbosity turn out to be a single direction in activation space that you can dial down Can we steer reasoning toward brevity without retraining?. A decoding-only penalty on switching ideas mid-thought improves accuracy, so the capacity to finish a line of thought was already there and was being under-used Do reasoning models switch between ideas too frequently?. If a strategy can be dialed up by a vector, a prompt structure or a decoding rule, it was already in the model.

What training adds is often when to use a strategy rather than what the strategy is. One reading of the corpus says RL post-training teaches models when to deploy mechanisms that pre-training already supplied How should reasoning systems actually be architected?. The same thinking mode can hurt or help depending on training. In vanilla models it produces self-doubt, and after RL it becomes useful gap analysis Does extended thinking help or hurt model reasoning?. The source of the latent material is also traceable. Across 5 million pretraining documents, reasoning draws on broad procedural knowledge (how-to patterns seen across many sources), while factual recall depends on narrow memorization Does procedural knowledge drive reasoning more than factual retrieval?. Under this picture a strategy is latent if some procedure like it appeared in pretraining, and the model recombines it.

The corpus offers two hints about what novelty would require. First, latent strategies stay bounded by their source: chain-of-thought degrades predictably when task, length or format shifts away from the training distribution, producing fluent reasoning whose logic doesn't hold Does chain-of-thought reasoning actually generalize beyond training data?. That looks like reuse of learned form, not invention. Second, reasoning models are 'wandering explorers'. They lack the validity, effectiveness and necessity that systematic search needs, so success drops exponentially as problems get deeper Why do reasoning LLMs fail at deeper problem solving?. A strategy that reliably meets those three properties would be a real capability gain rather than better elicitation. Creativity research adds another angle. It separates combinational, exploratory and transformational reasoning, and existing LLM methods address none of them Can LLMs reason creatively beyond conventional problem-solving?. My reading is that recombining known moves is the latent kind, and changing the space of moves is where genuine novelty would sit.

The corpus has no clean experiment that isolates a genuinely new strategy. The evidence is strong that elicitation explains most observed gains, and much thinner on what would prove a strategy was learned from scratch. The practical takeaway is that when a reasoning gain shows up, the first question is whether a steering vector or a better decoding rule could have produced it.