Merging code is a headache for every developer. Resolving conflicts, those nasty moments when Git throws its hands up, is like untangling a giant, digital knot. But what if AI could step in and handle this tedious task? Researchers are exploring exactly that, training large language models (LLMs) to automatically resolve merge conflicts. But how good are these AI-powered tools *really*? A new benchmark called CONGRA aims to find out. CONGRA introduces a clever way to categorize conflicts based on the complexity of the code changes, from simple text differences to complex functional modifications. By analyzing nearly 45,000 real-world conflict scenarios from popular projects like TensorFlow and the Linux Kernel, CONGRA provides a robust testing ground for automatic conflict resolution tools. The results are surprisingly nuanced. LLMs with massive contexts, which you’d expect to excel, don’t always outperform models with smaller contexts. Why? The researchers suspect that these ultra-large models haven’t been trained specifically on enough merge conflict data to fully leverage their capacity. Even more unexpected, general-purpose LLMs like Llama 3 sometimes outshine specialized code-focused LLMs. This highlights the importance of understanding not just the code, but also the surrounding semantics like comments and variable names—something that general LLMs, with their broader training, might be better equipped to handle. CONGRA's findings provide crucial insights for the next generation of AI-powered merging tools. While context seems important, researchers need to investigate how to best select and utilize the most relevant information around a conflict to maximize its impact. CONGRA is more than just a benchmark; it's a roadmap toward a future where merging code is smooth, automated, and maybe even… enjoyable.