Before an AI learns which steps to take, it first has to get each step right — is that how reasoning appears?

This explores whether the two-stage pattern seen in RL training, mechanics first and strategy second, explains how reasoning appears in language models. The clearest evidence in the corpus is a study across eight models. Early in RL training, learning is driven by execution correctness, meaning getting each step of a calculation right. Later, the bottleneck shifts to strategic planning. Uncertainty over planning tokens rises while uncertainty over execution tokens settles, and focusing optimization on the planning tokens gave significant gains Does RL training follow a predictable two-phase learning sequence?. On this view, reasoning "emerging" is less a sudden spark than a handoff. The model first becomes reliable at carrying out steps, then the limiting factor becomes choosing which steps to take.

The first phase needs raw material to consolidate. An analysis of 5 million pretraining documents found that reasoning leans on broad procedural knowledge, the how-to patterns spread across many sources. Factual recall depends on narrow memorization instead Does procedural knowledge drive reasoning more than factual retrieval?. Another cluster of notes argues the capability is already inside base models. RL steering, critique fine-tuning, decoding changes, SAE feature steering and RLVR all elicit reasoning that was already present, which points to elicitation being the bottleneck rather than capability acquisition Do base models already contain hidden reasoning ability?. Read together, phase one looks like turning procedures the model already half-knows into dependable execution.

Phase two connects to a stranger claim, that RL post-training teaches when to reason, not how. Hybrid models recovered 91% of the performance gains just by routing tokens, and activation vectors for reasoning strategies exist before any RL Does RL post-training create reasoning or just deploy it?. If the second phase is about strategy, that fits: the model is learning when to deploy what it has. Two neighboring findings fit too. Reasoning models often abandon promising lines of thought too early, and penalizing thought-switching tokens at decoding time improves accuracy with no retraining Do reasoning models switch between ideas too frequently?. Reasoning verbosity turns out to be a single steerable direction in activation space Can we steer reasoning toward brevity without retraining?. Both suggest the strategic layer can be adjusted from the outside. The corpus never tests the link between the second phase and deployment timing directly, so that connection is my reading of these notes, not something any one of them shows.

Two complications keep the picture from being tidy. First, the phases aren't tied to post-training. RLP treats chain-of-thought as an exploratory action during pretraining itself and gets about a 19% lift Can chain-of-thought reasoning be learned during pretraining itself?. Looped models such as Ouro do their reasoning during pretraining through iterative latent loops Can reasoning be learned during pretraining rather than after?. Second, a two-phase curve may describe a particular setup more than a law of nature. The reusable unit of reasoning training is a feedback interface entangled with the verifier, base model, optimizer, scaffold and budget, and changing any one of them changes what the same data does What is the actual reusable unit of reasoning data?. Token-level entropy is also a surface signal, and other work argues reasoning mostly happens in hidden-state trajectories, with visible text as a partial interface Where does LLM reasoning actually happen during generation?. The two-phase story is best treated as a useful map of what RL improves and when, not a full account of where reasoning comes from.