Imagine you're in a classroom, and the teacher poses a fiendishly difficult math problem. After a few minutes, most of the students raise their hands, all confidently presenting the same incorrect answer. They’ve all made the same subtle mistake in their logic. But in the back corner, one quiet student has worked through the problem differently and arrived at the correct, non-obvious solution. If you were to simply go with a show of hands—a majority vote—you’d walk away with the wrong answer. To find the truth, you’d need a way to recognize the brilliant, correct reasoning, even if it’s unpopular. This is the core challenge that Large Language Models (LLMs) face when solving complex reasoning tasks. A widely used technique to improve performance is to have the model generate multiple solutions and then select the most frequently occurring answer—what’s known as self-consistency or majority voting. This approach often works well for straightforward problems, but it breaks down when the problem is difficult enough that the majority of generated answers are wrong, yet all stem from the same flawed reasoning. This is where Meta’s new approach, AggLM, comes in. AggLM is a novel reinforcement learning method designed to go beyond simple majority rule. Instead of selecting the most common answer, AggLM evaluates the quality of each reasoning path—not just the final answer—by identifying which solutions are logically sound, well-structured, and aligned with correct principles. AggLM works by training a separate “judge” model to assess the reasoning steps of each generated solution. Rather than just counting how many times an answer appears, it weighs the validity of the logic behind each response. This allows the system to recognize when a minority of answers, though less frequent, are actually correct and well-reasoned—just like that quiet student in the back of the class. The method is especially powerful for tasks involving multi-step reasoning, such as solving advanced math problems, scientific reasoning, or complex code generation. By focusing on the quality of reasoning rather than popularity, AggLM reduces the risk of reinforcing systematic errors and improves accuracy on challenging problems. This shift represents a fundamental evolution in how we think about scaling LLMs. It moves away from the assumption that “more is better” and instead emphasizes “better reasoning is better.” AggLM doesn’t just make models smarter—it makes them more thoughtful. Meta’s work with AggLM underscores a broader trend in AI: the move from statistical patterns to deeper, more principled reasoning. As LLMs are increasingly tasked with real-world decisions—medical diagnoses, legal analysis, engineering design—the ability to find the right answer, not just the popular one, becomes critical. AggLM is a step toward building AI systems that don’t just mimic human thought, but actually reason like it.