Large language models (LLMs) have shown surprising abilities to solve problems, write stories, and even generate code. One trick that makes LLMs even better is "self-consistency": ask the LLM the *same* question multiple times, get slightly different answers each time, then pick the most common answer. This approach has worked wonders on shorter texts, boosting LLM performance significantly. But does this clever trick still hold up when LLMs face REALLY long documents? A new study investigated this question, examining how self-consistency fares on long-context problems where LLMs often struggle with "position bias"—favoring information at the beginning or end of a text and ignoring the middle. The researchers tested several LLMs on lengthy question-answering and text-retrieval tasks using datasets designed to mimic real-world scenarios like searching through long Wikipedia articles. Surprisingly, they discovered self-consistency doesn't help much with long texts. Not only does it fail to solve the position bias problem, in some cases it even *worsens* performance, particularly when pinpointing crucial information within a massive document. While larger LLMs generally performed better, the limitations of self-consistency persisted, suggesting this isn't just a matter of needing bigger models. The study also explored different prompt formats—how the question and document are presented to the LLM—and different parameters for self-consistency itself, but these tweaks only offered minor improvements. The researchers conclude that while self-consistency is a useful tool for shorter texts, it doesn't address the core issues LLMs face with long documents. They suggest future research focus on fundamentally rethinking how LLMs handle long contexts, potentially through specialized architectures or training methods that explicitly account for position bias. This might involve smarter ways to combine the multiple answers generated by self-consistency or even developing entirely new methods to overcome the limitations of current LLM technology. The challenge remains: how can we make AI truly understand and effectively use the wealth of information buried within extremely long texts?