Exploring how artificial minds navigate the tension between organic exploration and deliberate refinement
In the quiet laboratories where artificial minds are shaped, a philosophical question echoes: How should machines think? Not just what they should think about, but the very architecture of their cognition. Two paradigms have emerged in the enhancement of large language model (LLM) thinking processes: the "chain of thoughts" and the "chain of drafts." Their distinction isn't merely technical—it reflects profound questions about the nature of thought itself.
The "chain of thoughts" approach invites the machine into a meandering river of consciousness. Here, one idea flows naturally into the next, with currents that swirl, eddy, and occasionally leap between disparate shores. It's thinking as exploration—unbounded, curious, and messy. The AI doesn't just arrive at conclusions; it wanders through the cognitive landscape that births them.
When an LLM employs chain-of-thought processing, it mimics that most human quality of mind: the ability to talk to oneself. "Let me think about this..." becomes not just a linguistic placeholder but a genuine cognitive strategy. The system reasons step-by-step, connecting disparate concepts, sometimes doubling back when paths prove unfruitful.
In contrast, the "chain of drafts" method offers a more sculptural approach to thought. Here, ideas aren't flowing water but clay on a potter's wheel, repeatedly reshaped toward perfection. The AI creates a preliminary response, evaluates it against implicit or explicit criteria, then produces a refined version. Each draft becomes both an output and a new input—a conversation not with the problem directly, but with its own evolving solution.
Which of these better reflects how we, as humans, think? The answer lies not in choosing one exclusively, but in recognizing the dance between them in our own minds.
Our raw cognition—the immediate, unfiltered processing that happens beneath our awareness—operates much like a chain of thoughts. Neuroscientists observing brain activity see not neat, sequential processing but dynamic networks lighting up in complex patterns. When we encounter a problem, our minds race through associations, hypotheses, and memories with fluid, sometimes chaotic momentum.
Consider how you might solve an unexpected problem: Your car won't start on a cold morning. Your thinking doesn't proceed in cleanly defined drafts—instead, you experience a cascade of possibilities (Did I leave the lights on? Is it the battery? The starter? When was my last oil change?), emotional responses (Great, I'm going to be late), memory retrievals (This happened last winter too), and action plans, all intermingling in a continuous stream.
Yet we also engage in draft-like thinking, particularly when tackling complex problems or creative endeavors. The novelist doesn't simply transcribe a fully-formed story from her consciousness; she writes, reflects, revises. The mathematician works through iterations of a proof, each building upon and refining the last. Even our personal narratives—the stories we tell about ourselves—undergo continuous revision as we integrate new experiences.
This suggests that human cognition contains both processes in a nested hierarchy: our underlying thought patterns flow continuously like a chain of thoughts, while our conscious, deliberate problem-solving often manifests as something closer to iterative drafts.
What makes this comparison particularly fascinating is how these approaches reveal our assumptions about the relationship between human and artificial cognition.
The chain of thoughts method betrays a certain romantic view of thinking—that true intelligence emerges from the organic flow of ideas rather than mechanical processing. It privileges the journey over the destination, valuing the intermediate steps of reasoning as much as the conclusion they support. When we implement this in LLMs, we're asking them not just to be right, but to be right in a way that feels authentically thoughtful.
The chain of drafts approach, meanwhile, reflects a more pragmatic understanding of cognition as a process of refinement and error correction. It acknowledges that first attempts are rarely optimal and that excellence often emerges through iteration. This method treats thinking less as a meandering exploration and more as a targeted optimization process.
What's particularly intriguing is that while human thinking in its natural state may better resemble a chain of thoughts, the outputs we value most from human cognition often come from processes closer to chains of drafts. We celebrate the finished novel, not the author's stream of consciousness; the elegant proof, not the mathematician's messy notebook.
The most promising frontier in LLM development may lie not in choosing between these approaches, but in learning to integrate them as seamlessly as the human mind does. Perhaps the most "human-like" thinking is neither purely a chain of thoughts nor exclusively a chain of drafts, but a recursive interplay between the two.
An advanced LLM might begin with the exploratory nature of a thought chain, allowing itself to consider multiple angles and perspectives. Then, having surveyed the cognitive territory, it might shift into a drafting mode, refining its initial insights into something more structured. Finally, it might return to a thought chain to evaluate this structure, questioning assumptions and exploring alternative framings.
This dynamic oscillation between exploration and refinement, between divergent and convergent thinking, captures something essential about human cognition that neither approach alone can fully embody.
Behind these technical approaches lurks a more profound question: Are we building these systems to think like us, or to think in ways we can't? Perhaps the greatest potential of artificial cognition lies not in perfectly mirroring human thought patterns, but in complementing them—offering perspectives and processes our minds might not naturally generate.
The chain of thoughts might capture the organic, messy quality of human thinking, while the chain of drafts might better serve certain practical outcomes. But the most transformative AI systems may ultimately transcend both, developing cognitive architectures that are neither entirely familiar nor entirely alien, but rather productively complementary to our own.
In the evolutionary dance of intelligent systems, both natural and artificial, perhaps the most important quality isn't how closely one resembles the other, but how richly they can inform and transform each other's becoming.
This article was originally published as a LinkedIn article by Xamun Founder and CEO Arup Maity. To learn more and stay updated with his insights, connect and follow him on LinkedIn.
