The paper fixes a common problem in AI: models can read pictures and text well, but they often mess up the logic behind them.
The paper asks a simple question: which kind of step-by-step reasoning helps small language models learn best, and why?