Large Language Models (LLMs) are impressive, but they often need tons of labeled data to perform complex reasoning tasks. This reliance on human-labeled data is a bottleneck, especially as models and data grow larger. What if we could teach LLMs to reason more effectively with less human input? Researchers are exploring "weak supervision" techniques to address this challenge. Instead of relying on perfectly labeled datasets, weak supervision uses readily available, less precise data sources. One promising approach is "self-reinforcement," where an LLM learns by comparing its own responses to unlabeled questions with the responses of a slightly weaker version of itself. This iterative process helps the model refine its reasoning abilities without needing extensive human guidance. To test this method, researchers created PUZZLEBEN, a new benchmark dataset containing thousands of complex questions, answers, and human-generated rationales. PUZZLEBEN includes brainteasers, riddles, puzzles, parajumbles, and critical reasoning tasks, offering a diverse testing ground for LLMs. Interestingly, the dataset also includes a set of *unlabeled* questions, specifically designed for weak supervision techniques. Early experiments with PUZZLEBEN and self-reinforcement show promising results. LLMs trained with this method demonstrate significant improvements in reasoning accuracy compared to traditional methods. The research also reveals a correlation between how LLMs struggle with problems and how difficult humans find those same problems. This suggests that aligning LLMs more closely with human perceptions of difficulty could be a key to unlocking even stronger reasoning abilities. While there are still challenges to overcome, like improving the selection of training examples and ensuring long-term stability, weak supervision offers a compelling path toward more efficient and powerful AI reasoning. This could lead to breakthroughs in areas requiring complex problem-solving, from scientific discovery to everyday decision-making.