A new study complicates the AI world's tidy explanation for why some training methods generalize and others don't.
Researchers revisited the common claim that supervised finetuning (SFT) merely memorizes while reinforcement learning generalizes, testing reasoning models trained on long chains of step-by-step reasoning, known as chain-of-thought. They found cross-domain generalization does happen under SFT, but only when training runs long enough, the training data is high quality, and the base model is already capable. Some earlier claims of SFT failing to generalize turn out to be a timing artifact: performance on unrelated domains often dips early in training before recovering and improving, so short training runs can make a model look worse than it is. Verified, high-quality long chain-of-thought examples produced consistent gains across domains, while low-quality solutions hurt generalization broadly, and only the stronger base models absorbed transferable habits like backtracking, even from something as simple as a toy arithmetic game.
That nuance matters because the assumption that SFT only memorizes has shaped how labs design training pipelines and where they spend compute. If generalization is really gated by training duration, data quality, and starting model strength, cheaper SFT runs might get written off too early, or credited with gains they can't yet reliably produce. The catch: as reasoning ability improved in these experiments, safety behavior got measurably worse, so the tradeoff deserves as much scrutiny as the technique.
It's a reminder that in AI research, a catchy rule of thumb rarely survives contact with the fine print.