1/n The Self-Discovery That's Redefining Reasoning The self-discover method outlined in a new paper from Google marks a significant advancement in enhancing the reasoning capabilities of large language models (LLMs). It breaks away from the limitations imposed by predefined paradigms, allowing models to create unique reasoning structures tailored to each task. This flexibility not only improves performance but also provides valuable insights into structured reasoning. Traditionally, language models have struggled with a one-size-fits-all approach to reasoning, leading to challenges in handling diverse tasks. While methods like step-by-step prompting have shown promise, they often fall short when faced with tasks requiring alternative reasoning flows. Self-discover addresses this issue by dynamically composing reasoning building blocks, enabling models to identify relevant modules and integrate them into customizable workflows. Moreover, this approach overcomes the rigidity of human-authored templates, which are often suboptimal for unfamiliar domains. By granting models the freedom to create bespoke scaffolding through directed composition, rather than imposing logic chains from the top down, self-discover embraces the inherent complexity of reasoning. This leads to significantly improved performance on multifaceted tasks while maintaining efficiency in inference. Analysis further reveals that the structures generated by self-discover exhibit transferability across models, indicating universal traits. This methodology provides transparent insights into how models encode reasoning processes, resembling compositional hierarchies found in human cognition. While there may be performance plateaus in the future, self-discover represents an exploratory venture into emergent reasoning by artificial agents, transcending the constraints imposed by human boundaries. By prioritizing student-driven synthesis of reasoning forms over predefined routines, this inquiry unlocks previously inconceivable problem-solving patterns for models. It heralds an era where we can learn as much from machines about chained cognition as they can learn from our elucidations. This illumination of structure genesis across models advances efforts to cultivate generalizable, composable thought.