Instead of teaching an AI to reason after training, what if you built it in by looping its layers?

This explores whether building reasoning into a model's architecture and pretraining, by looping the same layers repeatedly, beats teaching it afterward with supervised fine-tuning (SFT). The corpus has no head-to-head test of the two. The separate pieces do point the same way, though. Looping looks better on the quality of reasoning. On cost, the evidence covers parameter count but not compute.

The looped side is strongest on parameter efficiency. Ouro's 1.4B–2.6B models match 12B baselines by doing reasoning inside the pretraining loop, not by storing more knowledge (Can reasoning be learned during pretraining rather than after?). The wider looped-model work explains why. Re-applying layers gives a model state tracking and compositional generalization that adding parameters doesn't, and convergence of the loop gives a natural signal for when to stop (Can models learn by looping instead of growing larger?). A side benefit is that Ouro's intermediate latent states line up with its final outputs. That makes its reasoning more faithful than a chain-of-thought trace that can drift from the answer.

SFT's problem is that it can raise scores without building the skill. It improves final-answer accuracy while cutting Information Gain per reasoning step by 38.9 percent, so models get to correct answers through post-hoc rationalization instead of real inferential steps (Does supervised fine-tuning improve reasoning or just answers?). Three separate tests back this up. After fine-tuning, cutting the reasoning short, paraphrasing it, or swapping in filler leaves the answer unchanged more often, which means the written reasoning is decoration (Does fine-tuning disconnect reasoning steps from final answers?). Chain-of-thought skills learned this way also fall apart predictably once tasks shift in length, format, or type. The result is fluent reasoning-shaped text without valid logic behind it (Does chain-of-thought reasoning actually generalize beyond training data?).

The comparison may be slightly off, because most of the evidence says post-training isn't where reasoning is created. Base models already hold latent reasoning that five different methods can elicit (Do base models already contain hidden reasoning ability?). RL post-training mostly teaches a model when to deploy reasoning, not how (Does RL post-training create reasoning or just deploy it?). If that's right, SFT is a weak way to add capability, and the real contest is over what happens during pretraining. Looping is one route. Rewarding chain-of-thought during pretraining is another, and it lifted math and science benchmarks by about 19 percent (Can chain-of-thought reasoning be learned during pretraining itself?). Analysis of 5 million pretraining documents adds that reasoning generalizes from broad procedural knowledge in the pretraining data, unlike facts, which need memorization (Does procedural knowledge drive reasoning more than factual retrieval?). Even within post-training, SFT isn't the strongest option. Reinforcement learning that rewards explanation quality embeds domain knowledge better than SFT does (Can reinforcement learning embed domain knowledge more effectively than supervised fine-tuning?).

The corpus doesn't say whether looping is cheaper in compute. Ouro's win is measured in parameters, and every extra loop costs another pass through the layers. Nothing here compares total training or inference cost against an SFT pipeline. What the corpus does support is narrower. Looping builds reasoning that is more faithful and less dependent on model size, while SFT tends to buy benchmark points and leaves the reasoning steps less connected to the answers.