Key Aspects of the Research:

The researchers tested how 23 modern LLMs handle one of the toughest exams for financial analysts (CFA Level III), using different prompting approaches. They found that the best models can pass the exam, but only with advanced prompting techniques such as chain of thought ("think step by step"), which especially boosts the quality of answers to complex essay questions. A plain, head-on request (zero-shot) gives noticeably worse results.

Key Result: Advanced prompting techniques such as chain of thought are critical for solving complex analytical tasks, turning an LLM into a professional-grade tool.

The Core of the Method, Explained:

The core of the method the research proves effective is not just asking the model a question but steering its thinking process. Instead of expecting an instant, correct answer to a complex request, we make the model first build a logical chain of reasoning and only then give the final answer.

The research highlights two main advanced approaches:

1. Chain-of-Thought (CoT): This is the most practical and powerful technique. Right in the prompt, you instruct the model to work in steps.

   • Step 1: Decomposition. Ask the model to first restate the task and break it into parts.
   • Step 2: Reasoning. Tell it to "think step by step," analyzing each aspect of the problem and showing its calculations or logic.
   • Step 3: Synthesis. At the end, ask it to pull all the reasoning together and give a final, well-justified answer. This approach keeps the LLM from "hallucinating" the first answer that comes to mind and makes it build the answer on logic, which sharply improves quality and accuracy.

2. Self-Discover: This is a more advanced version of CoT. Here you ask the model to first plan its work.

   • Step 1: Choosing the tools. The model has to decide which "reasoning modules" it will need (for example: "data analysis," "comparing options," "applying a rule").
   • Step 2: Building a plan. The model has to draw up a plan for how it will apply those modules to the specific task.
   • Step 3: Executing the plan. The model follows its own plan to generate the answer. This method turns the LLM into its own "project manager," which is especially useful for multi-part and creative tasks.

The main takeaway for users: For any task more complex than a simple factual question, the prompt should spell out not only WHAT to do but HOW to do it, guiding the model's thinking process.

Practical Applicability:

• Direct applicability: High. You can start using chain of thought right away. Just add phrases like "Think step by step," "Break the task into stages," "First make a plan, then write the answer" to your prompt. The templates in the paper's appendix (Fig. 7 and 8) can be adapted directly to any task by swapping the financial context for your own.

• Conceptual value: Very high. The research clearly proves that an LLM isn't just a "black box." Answer quality depends directly on how we direct the model's "attention" and "logic." It gives you a "mental model" of the LLM as a tool you configure with instructions rather than an all-knowing oracle. The idea of a quality/speed/cost trade-off is also key to understanding how LLMs are used in practice.

• Adaptation potential: Maximum. Although the research focuses on finance, the techniques it describes (CoT, Self-Discover) are completely universal. "Breaking a task into parts," "step-by-step reasoning" and "planning ahead" are meta-skills of problem solving that apply to writing a marketing strategy, planning a trip, creating a video script or analyzing a legal document. Adapting them comes down to swapping the subject area in the prompt template.

A Practical Example:

Here's an example prompt that uses chain of thought (CoT) for an everyday business task: analyzing customer reviews and developing proposals.

You are an experienced marketing analyst at a company that makes and sells organic food. Your task is to analyze customer reviews and propose improvements to the marketing strategy.

**Context (customer reviews):**
1.  "I really love your yogurts, the taste is so natural! But the packaging is awkward, the lid keeps tearing."
2.  "The price of your granola seems inflated compared to competitors. What exactly am I paying extra for?"
3.  "Found you by accident in a small health food store. Why aren't you in the big supermarkets? It would be easier to buy."
4.  "I follow your Instagram, but it's just product photos. I'd love to see recipes and learn more about the farmers who supply your milk."

**Assignment:**
Analyze these reviews and propose 3 key, practical ideas for improving the marketing strategy.

**How to complete it (chain of thought):**
Follow these steps to structure your answer:

1.  **Restate the task:** Briefly describe in your own words what is required of you.
2.  **Identify the key issues:** Analyze each review and group the customers' complaints and suggestions by main theme (for example, "Packaging," "Price," "Distribution," "Content").
3.  **Outline your answer:** Create a short 3-point plan where each point is an idea you're proposing.
4.  **Work out the details:** Following your plan, describe each of the three ideas in detail. Explain which customer problem each idea solves and what concrete steps are needed to implement it.
5.  **Final summary:** At the end, briefly sum up your three proposals.

**Get started:**

Why This Works:

This prompt is effective because it puts the chain of thought principles described in the research into practice:

1. Structured thinking: Instead of generating ideas right away, the model has to follow a clear algorithm: analysis -> grouping issues -> planning -> detailing. That prevents shallow or irrelevant answers.
2. Task decomposition: The instruction "Identify the key issues" makes the model first process and organize the input (the reviews) instead of trying to solve everything at once. This lowers the cognitive load and improves the accuracy of the analysis.
3. Forced justification: The "Work out the details" step requires the model not just to name an idea but to explain why it's good ("which problem it solves") and how to implement it ("concrete steps"). That directly increases the depth and practical value of the answer.
4. Role: Assigning the role of "experienced marketing analyst" tunes the model to the right vocabulary and mindset, which further improves the answer.

Another Practical Example

Here's an example prompt that uses the Self-Discover method to build a content plan for a personal blog.

You are an expert in content strategy and SEO. Your task is to develop a content plan for a blog for the next month.

**Blog context:**
*   **Topic:** Road trips across the US.
*   **Target audience:** Families with kids (ages 30-45) looking for ideas for short weekend getaways and vacations.
*   **Blog goal:** Become a go-to resource, grow the audience and eventually monetize through affiliate programs (hotels, car rentals).

**Assignment:**
Create a detailed content plan for 4 weeks (8 posts, 2 per week).

**How to complete it (Self-Discover):**
Before writing the content plan itself, develop a reasoning structure.

**1. Choose reasoning modules:** Identify and list the key thinking modules needed to create a high-quality content plan. (Examples: Audience analysis, Topic generation, SEO optimization, Content formatting, Calendar planning).

**2. Adapt the modules to the task:** For each module you chose, briefly describe how exactly you'll apply it to this particular travel blog.
    *   *For example, for "Audience analysis":* "I'll focus on the pain points and interests of families with kids: safety on the road, entertainment for kids, budget-friendly lodging and food."

**3. Build the content plan structure:** Based on the previous steps, create a high-level outline for your answer.

**4. Execute the plan and write the content plan:** Following your structure, write the full content plan. For each of the 8 posts, specify:
    *   Working title (a catchy headline).
    *   SEO keywords.
    *   Short description (2-3 sentences on what the post is about).
    *   Format (for example: "guide article," "top-5 list," "interview," "photo essay").

**Start by developing the reasoning plan:**

Why This Example Works: The Mechanism.

This prompt uses the Self-Discover methodology, making the model engage in metacognition: thinking about how it will think.

1. Forced planning: Steps 1 and 2 ("Choose modules" and "Adapt modules") keep the model from rushing straight into generating post ideas. Instead, it has to build a framework for its work first. This ensures every important aspect (audience, SEO, formats) gets covered.
2. Building its own logic: The model doesn't just follow someone else's steps (as in CoT); it decides for itself which "thinking tools" it needs. This is especially effective for complex, multifaceted tasks like strategic planning, where there's no single right path.
3. Greater relevance: Adapting the modules to the specific task (step 2) makes the model dig deep into the context (a blog about family road trips) and ensures the final content plan precisely matches the goals and audience instead of being a pile of generic travel ideas.
4. Structured result: The step 4 requirement to give each post a title, SEO keywords, a description and a format produces a ready-to-use, well-organized document rather than a continuous stream of text.

Usefulness Score: 92