LLMs are underutilized due to sub optimal management (alexzhang13.github.io)

Recent insights suggest that frontier language models (LMs) in AI may be significantly underutilized due to inadequate management techniques rather than inherent limitations in their capabilities. The "mismanaged geniuses hypothesis" (MGH) posits that existing LMs are exceptionally powerful at solving a wide array of tasks but struggle with long-horizon and iterative reasoning, often due to the rigid, human-engineered frameworks that define how these models decompose and tackle problems. Instead of continuing the trend of merely scaling model size and data, the research advocates for equipping LMs with the ability to self-manage by natively decomposing tasks into subtasks, which may unlock greater potential and efficiency in tackling complex issues. Key advancements, such as coding agents and recursive language models (RLMs), highlight promising pathways for enabling LMs to effectively manage their own workflows. These systems allow LMs to generate intuitive plans for completing tasks, thus facilitating smoother task decomposition and enhanced performance on long-horizon challenges. By exploring the space of permissible decompositions and training LMs to leverage these structures, researchers believe that LMs can solve out-of-distribution problems more efficiently. The implications for the AI/ML community are profound: rather than fixating solely on model size, a shift toward better task management could lead to significant advancements in AI capabilities.