The world of information is no longer confined to single sources. We gather insights from multiple news articles, research papers, and user reviews to form comprehensive understandings. This multi-source reality poses a unique challenge for Large Language Models (LLMs). How can these powerful AI tools effectively synthesize information from diverse, sometimes conflicting, sources? A new benchmark called SEAM (Stochastic Evaluation Approach for Multi-document tasks) aims to answer this question. SEAM tests LLMs on various multi-document tasks, including summarization, question answering, and coreference resolution—essential skills for navigating our complex information landscape. What sets SEAM apart is its focus on real-world scenarios. It acknowledges the messiness of information, incorporating factors like conflicting reports and the lack of inherent order in document collections. By testing LLMs under these conditions, SEAM provides a robust evaluation of their ability to handle multi-source input. The research reveals that even state-of-the-art LLMs struggle with these tasks. Surprisingly, simply increasing model size doesn’t guarantee better performance. The challenge isn't just about processing longer texts; it's about synthesizing and reconciling information from different perspectives. One key takeaway is the importance of consistent evaluation methodologies. LLMs are sensitive to even minor changes in prompts and formatting. SEAM’s stochastic approach, which involves repeated evaluations with varied prompts, offers a more reliable and robust evaluation method. This research is crucial for understanding the current limitations of LLMs and guiding future development. By developing benchmarks like SEAM, we can push the boundaries of AI, creating models capable of truly understanding and synthesizing the complexities of multi-source information.