Large language models (LLMs) are getting remarkably good at answering complex scientific questions. But what happens when the questions themselves are flawed? A new research project called SciFaultyQA explores this intriguing problem, revealing that LLMs often struggle to identify nonsensical or illogical science questions, and sometimes even attempt to answer them. Researchers found that LLMs like GPT-4 and Gemini frequently miss faulty reasoning in questions, highlighting a potential blind spot in their abilities. To investigate this further, the team developed a clever, GAN-inspired method to create a dataset of intentionally flawed science questions. This involved using different LLMs to generate faulty questions and another LLM to act as a “discriminator,” trying to spot the errors. This iterative process helped create a challenging benchmark called SciFaultyQA to test how well LLMs can detect these tricky questions. The results showed that even advanced LLMs have a surprisingly low success rate in identifying these faulty questions. However, the researchers discovered that giving LLMs access to tools like internet search dramatically improved their ability to spot the flaws. Integrating multiple LLMs into a multi-agent system where they check each other's work also boosted performance. This research underscores the importance of developing robust methods to evaluate and improve LLMs' critical thinking skills. It also suggests that augmenting LLMs with external tools and collaborative frameworks can be a powerful way to enhance their reasoning abilities and make them more reliable in real-world applications.