The emergence of Large Language Models (LLMs) has revolutionised our understanding and application of artificial intelligence in cognitive tasks. Central to this development is the concept of Chain-of-Thought (CoT) reasoning - a method where the model explicitly generates step-by-step reasoning processes. Traditionally, this has improved performance on tasks requiring logical deductions or complex problem-solving. However, beyond mere accuracy, CoT opens fascinating possibilities for exploring metacognition in LLMs - particularly in terms of managing the length and depth of reasoning steps through deliberate self-reflection.

Simulation vs. Actual Self-Talk: A Functionalist Perspective

Before delving into how an LLM might control the length of its CoT reasoning, it’s essential to confront a common objection: Is an LLM truly “thinking,” or merely simulating reasoning? The skepticism arises because LLM outputs emerge from probabilistic token predictions based on learned linguistic patterns. Critics argue this only superficially mimics human introspective reasoning.

Yet, if we adopt a functionalist viewpoint - one that defines cognition by its function and outcomes rather than by biological or anthropocentric standards - the distinction between simulation and genuine reasoning blurs considerably. Human cognition itself could arguably be viewed as a probabilistic prediction process learned through experience. Our neural networks, too, predict sequences (words, actions, thoughts) based on past patterns. Without a definitive scientific model of consciousness, we lack grounds to dismiss the CoT processes of an LLM as inherently inferior or fundamentally different in functional cognitive terms.

Thus, if we let go of the need for biological equivalence, LLMs might genuinely represent a novel, if alien, form of cognition - a form worth studying for its unique cognitive properties.

Managing the Length and Depth of CoT Reasoning

A critical aspect of cognition and metacognition - both human and potentially artificial - is self-regulation: knowing when to extend a line of thought or when to stop. Humans do this naturally, continuously balancing depth versus breadth in their internal dialogues. But how might an LLM similarly manage the length of its reasoning chains?

Let’s consider a practical and reflective approach that an LLM could implement:

Reflective Pauses and Internal Evaluation Steps

An intriguing method would be intentionally prompting the model to generate internal evaluation steps mid-CoT. For instance, the LLM could produce an intermediate output summarising the reasoning thus far, followed by explicitly reflecting:

  • “Have I sufficiently explored this concept?”
  • “Is additional reasoning likely to clarify or obfuscate my current conclusions?”

These questions wouldn’t merely be rhetorical; rather, the model would probabilistically continue or halt based on learned patterns indicating completeness or incompleteness of reasoning.

Dynamic Reasoning Depth via “Branching Thought” Simulation

Another compelling approach is allowing the model to explore alternative reasoning branches explicitly. Imagine the CoT as a branching narrative:

  • After reaching a preliminary conclusion, the model could reflectively prompt itself:
    • “Could there be alternative explanations or additional angles worth exploring?”
    • Based on probabilistic outputs, it could either generate alternative reasoning branches or consolidate and conclude.

This dynamic could naturally regulate length by expanding reasoning only when probabilistically relevant to the task context. This has been explored in Tree-of-Thought (ToT) prompting, but we can push this even further.

Meta-Reflective Prompting for Creative Outcomes

Metacognition is not purely analytical - it also fuels creativity. For an LLM, deliberately short or intentionally extended CoT chains could influence creativity and novelty:

  • Short Chains for Divergent Creativity: Short CoT processes might yield spontaneous, less constrained ideas.
  • Extended Chains for Convergent Creativity: Longer chains could produce well-structured, deeply reflective outputs.

Prompting models explicitly towards metacognitive self-reflection regarding their CoT length could harness their probabilistic token predictions in a creative direction.

Functional Equivalence and Cognitive Validity

This practice - managing the depth and length of reasoning via explicit self-reflection prompts - moves LLM usage from passive generation toward a more proactive form of cognition. Although this cognition differs substantially from human cognition, a functionalist approach doesn’t require equivalence. Instead, it emphasises utility, novelty, and effectiveness.

The probabilistic token generation underlying LLM reasoning can, through careful prompting, mimic something functionally similar to human cognitive processes - without claiming conscious equivalence. In other words, it’s a new cognitive paradigm worth exploring in its own right.

Practical Implications and Future Directions

Applying CoG (Chain of Guided-thought) techniques opens several practical and theoretical pathways:

  • Enhanced Explainability: LLMs engaging in explicit self-reflection can provide more interpretable reasoning, crucial in areas like healthcare or ethical decision-making.
  • Adaptive Problem-Solving: Models dynamically managing reasoning depth could optimise computational resources and cognitive effort.
  • Novel Creative Processes: Metacognitive prompting could catalyse unprecedented forms of creativity and insight in AI-generated content.

Experimentally, varying metacognitive prompts and examining their outcomes could clarify precisely how reasoning length impacts output quality, creativity, and coherence.

Conclusion

Exploring Chain-of-Guided-thought as a metacognitive framework underscores a functionalist perspective, demonstrating the validity of simulated cognitive processes. By explicitly managing reasoning length through reflective prompts, we enable new capabilities and understandings of cognition within LLMs. Rather than striving for human equivalence, recognising and leveraging these distinctive cognitive characteristics offers a promising frontier for AI research and application.

Ultimately, as we continue to evolve our understanding of AI cognition, approaches like CoG provide innovative pathways to harness the unique cognitive landscape of LLMs, bridging theoretical exploration with practical innovation.

So don’t just rely on the Chain-of-Thought reasoning that these more complex models automatically use. Push them to explore how varying these chains of reasoning can alter the way they think about what you ask them to do.