Have you tried these reasoning techniques with Copilot?

I have spent the last few months working closely with Microsoft and some of their largest customers as they seek to implement Copilot (Office 365) into their organizations. I'll cover implementation and secure connection of enterprise data sources (i.e. non-Office 365) to Copilot in another post, but for this post I'll focus on how your approach to prompting Copilot can drastically change your experience.

In order to do this, let me introduce you to the concept of "reasoning techniques". Think of these as cognitive techniques to reach and end goal, each with different approaches as to how you would process information, reason, make decisions and determine your next actions based on previous information.

These reasoning techniques when applied to the field of agent cognitive architectures make use of prompt engineering and task planning.

Put more simply, the question you ask and the framework you provide the model to answer, will change the results that you see.

Let's see it in action and learn more about three of the most popular techniques: ReAct, Chain-of-Thought (CoT) and Tree-of-Thoughts (ToT).

ReAct (Reasoning + Acting) Prompting

ReAct is a paradigm that combines reasoning with action planning to enable LLMs to solve complex tasks through interleaved reasoning traces and actions. The key aspects of ReAct include:

• Generates both verbal reasoning traces and task-specific actions that can interact with external environments or tools
• Allows models to retrieve additional information through actions and use reasoning to determine what to retrieve next
• Helps overcome issues like hallucination and error propagation by enabling interaction with external sources

Example Use Case: Troubleshooting a computer issue

Prompt:

Expected Result:

Chain-of-Thought (CoT) Prompting

Chain-of-Thought is a prompting technique that breaks down complex problems into sequential reasoning steps. Its main characteristics are:

• Structures responses by guiding LLMs through intermediate steps toward a final answer
• Works best with larger models (≥100B parameters) as smaller models may produce illogical chains
• Particularly effective for tasks requiring multi-step reasoning like math problems and logical deduction

Example Use Case: Calculating total cost for a party budget

Example Prompt:

Expected Result:

Tree-of-Thought (ToT) Prompting

Tree-of-Thoughts extends CoT by enabling exploration of multiple reasoning paths simultaneously. Key features include:

• Maintains a tree structure where thoughts represent coherent language sequences serving as intermediate steps
• Allows for backtracking and exploring different branches when a particular path proves unsuccessful
• Combines generation and evaluation of thoughts with search algorithms like breadth-first and depth-first search

Example Use Case: Planning a weekend trip itinerary

Prompt:

Expected Result:

Key Differences

ReAct differs from CoT by incorporating actions and external knowledge sources, while ToT expands on CoT by exploring multiple potential solutions simultaneously rather than following a single chain of reasoning.

Summary

So, what have we learned? That there are multiple strategies on how to engineer your prompts, each with their own advantages and disadvantages depending on your desired use case.

Understanding these strategies is growing in importance as Copilot expands rapidly to Agentic AI through the use of Copilot Studio and Custom Engine Agents.

But like everything in life, there is a counter argument. As models and their workflows start to automatically add reasoning/deeper thinking steps the need for users to perform these steps at prompt time is reduced or in some cases mitigated. Abram Jackson (Microsoft Copilot Extensibility PM) covers this is great detail here.

Try the prompts above in your own Copilot to see them in action!