The ARC (AI2 Reasoning Challenge), a popular benchmark for testing the reasoning abilities of Large Language Models (LLMs), has been considered a significant hurdle for AI. However, new research suggests the challenge isn't as tough as we thought—the problem lies not in the questions themselves, but in how we've been asking them. Traditionally, LLMs were presented with each answer choice in isolation, preventing them from comparing options like a human would. This approach significantly hampered their performance, especially on questions requiring comparison, such as identifying the heaviest object or the best explanation. The research reveals that when LLMs see all answer options together, their accuracy on ARC Challenge skyrockets, closing the gap between AI and human performance dramatically. This isn’t just about ARC; similar effects have been observed in other benchmarks like OpenBookQA and SIQA. By simply presenting the questions in a more natural, comparative context, LLMs have demonstrated surprisingly strong reasoning abilities, even surpassing human performance in some cases. This highlights the importance of crafting evaluations that truly reflect real-world scenarios, ensuring we're measuring AI's actual capabilities, not its ability to overcome artificial constraints. While some researchers have quietly adopted this more effective evaluation method, the shift hasn't been universally acknowledged. This oversight underscores the need for greater transparency and standardization in LLM evaluation practices to avoid misinterpreting AI's potential. As AI continues to evolve, rethinking how we design and interpret benchmarks will be crucial for accurately gauging its progress and unlocking its full capacity.