Part 2/11:
Prior to this breakthrough, most large language models (LLMs) relied heavily on parallel thinking. In essence, the model would generate multiple potential solutions or reasoning paths in parallel and then select the most common answer—a process akin to polling or majority voting. While effective to a degree, this approach had significant limitations:
Diminishing Returns: After generating hundreds or thousands of solutions, further improvements in accuracy plateaued or even declined due to noisy or flawed reasoning paths polluting the results.
Cost and Efficiency: Producing numerous solution traces consumed vast amounts of tokens, resulting in increased computational costs, longer processing times, and inefficiencies.