Large language models (LLMs) possess a fascinating ability called "in-context learning"—the power to figure out rules from a few examples, just like humans do. But how do they actually *do* it? New research challenges the prevailing theory of a central "task vector" encoding the learned rule. Instead, when faced with tasks requiring multiple examples to grasp the underlying pattern, LLMs take a more distributed approach. Imagine learning to categorize objects: you see several items labeled "fruit" or "vegetable," gradually forming a mental boundary between the groups. LLMs behave similarly, storing rule information not in one place, but distributed across multiple "rule vectors." Each example contributes a small piece to the puzzle, forming a richer understanding of the overall concept. These rule vectors aren't simple memorizations either. They hold abstract representations of the relationship between inputs and outputs, allowing the LLM to apply the rule even when presented with completely new data. This distributed, abstract representation is critical for tasks that rely on drawing connections between multiple examples, rather than simple input-output mappings. For instance, in the categorization task of determining whether a string of characters is longer than 5, the LLM needs several examples to truly learn the pattern. The research demonstrates that these distributed rule vectors play a significant role in such scenarios, offering new insights into the inner workings of in-context learning. This discovery has big implications for the future of AI. By better understanding how LLMs learn from examples, we can develop better training methods to improve their reasoning abilities, especially in complex scenarios requiring understanding from multiple sources. This unlocks opportunities for LLMs to handle ever more sophisticated tasks, leading to more robust, versatile AI systems.