How does non-reasoning SFT prevent overfitting before RL training begins?

This reads the question as asking how an SFT stage, done before RL begins, keeps a model from overfitting. The honest synthesis: the collection mostly documents the opposite — SFT tends to *introduce* the overfitting that RL later has to fight. One careful study of SFT-then-RL pipelines finds a predictable three-phase arc — the model is first disrupted by the policy shift, readapts to the expert examples, then overfits to them Why does SFT-then-RL training follow a predictable three-phase pattern?. So a long SFT warm-up doesn't prevent overfitting; it walks the model into it.

What SFT actually teaches is part of the problem. Two notes show it buys surface competence at the cost of substance: fine-tuning raises final-answer accuracy while cutting reasoning informativeness by nearly 39%, with models reaching answers through pattern-matched shortcuts rather than genuine inference Does supervised fine-tuning actually improve reasoning quality?, and on optimization problems it makes outputs *look* right — clean JSON, valid structure — without making them physically feasible Does supervised fine-tuning actually improve reasoning on optimization problems?. That is overfitting to form. A 'non-reasoning' SFT pass that only teaches formatting and answer-matching is exactly the kind that locks in shortcuts.

So why do people run SFT first at all? The better framing in the corpus is that the reasoning is already there. Several independent lines of evidence find base models contain latent reasoning capability that minimal training merely unlocks Do base models already contain hidden reasoning ability?, and that RL post-training teaches a model *when* to reason rather than *how* — hybrid setups recover 91% of the gains just by routing tokens Does RL post-training create reasoning or just deploy it?. Under this view, a heavy reasoning-SFT stage risks overwriting capability the model already has; keeping SFT light and non-reasoning leaves the latent skills intact for RL to elicit rather than baking in imitation patterns.

The more compelling answer to overfitting isn't a clean SFT-then-RL handoff at all — it's dissolving the boundary. The same study that names the shift-readapt-overfit arc resolves it by folding SFT *into* on-policy RL as a dynamically weighted auxiliary objective, so the model never gets a chance to overfit to a frozen expert set before RL starts Why does SFT-then-RL training follow a predictable three-phase pattern?. RL has its own overfitting traps to watch — it collapses onto a single dominant pretraining format within the first epoch Does RL training collapse format diversity in pretrained models?, and overly hard samples breed degenerate shortcuts that contaminate existing skills Do overly hard RLVR samples actually harm model capabilities? — so the goal is less 'prevent overfitting with SFT' than 'don't let either stage memorize at the other's expense.'

The thing you might not have expected: the safest role for SFT before RL is to do *less*, not more. A minimal, formatting-only pass that preserves the base model's latent reasoning — rather than a thorough reasoning-imitation stage — is what keeps RL's elicitation room open.