Prompt Engineering: Intermediate Techniques

Andrew Ellis

28 November, 2024

Chain of Thought (CoT) prompting

Techniques: Encourage the model to proceed in a step-by-step manner. This has the effect of making the desired output more probable. The output looks like the LLM is showing its reasoning process¹.

Example Prompt

Think through this step-by-step: 1) List the symptoms 2) Consider possible causes 3) Evaluate urgency 4) Recommend action

Often it can be sufficient to just ask the model to think step-by-step.

Example Prompt

Think step-by-step.

1. This behaviour has been trained into recent models.

Why Chain of Thought?

• Amount of computation is constant per token.
• By forcing the LLM to generate more (useful) tokens, it will therefore generate more (useful) content.
• This in turn narrows the space of possible outputs, and steers the model towards regions of the output space that are more desirable.

Drawbacks of Chain of Thought

• LLM performance on reasoning problems does not generalize well
• Chain of thought prompting aims to mitigate this by demonstrating solution procedures
• Stechly, Valmeekam, and Kambhampati (2024) found meaningful performance improvements only with highly problem-specific prompts.

Few-Shot Learning

Technique: Provide multiple examples before asking for a new output.

The way that we structure Few-Shot Prompts is very important. By this, we mean do we separate the inputs and outputs with a colon (:) or the words INPUT/OUTPUT. We have seen examples of both earlier in this article. How can you decide? We generally use the input: output format and occasionally use the QA format, which is commonly used in research papers.

Use 2-5 examples for simple tasks. Use often ~10 examples for harder tasks

Example Prompt

Input: “Great product, 10/10”

Ouput: “Great product, 10/10”: {“label”: “positive”}

Input: “Didn’t work very well”

Output: “Didn’t work very well”: {“label”: “negative”}

Input: “Super helpful, worth it”

Output: “Super helpful, worth it”: {“label”: “positive”}

Input: “I’m not sure I would buy this again”

Output:

Structured Output

Technique: Specify a structure for the model’s response.

Example Prompt

Provide your assessment in JSON format:

{
  "severity": "[Emergency/Urgent/Non-urgent]",
  "potential_causes": "[List top 3]",
  "recommended_action": "[Specific next steps]"
}

Self-Consistency

LLMs are prone to variability in their responses.

Technique: Generate multiple answers, aggregate the responses and select the majority result.

Do this several times:

Example Prompt

Provide three independent assessments for these symptoms.

Think step-by-step.

Symptoms: [insert symptoms here]

Provide the responses to an LLM in a new session:

Example Prompt

Analyze whether the following assessments agree with each other. Give me your expert assessment based on the assessments you received.

Bonus tip: keep up with prompt engineering research

Technique: Use LLMs to “read” new research papers.

Example Prompt

Based on the attached research paper on [prompt engineering technique], write a prompt that would cause an LLM to behave according to the techniques described in this paper. Use [topic] as an example.

Problems with prompt engineering

• “Positive thinking” prompts have inconsistent effects across models.
• Chain of Thought (CoT) prompting generally improves performance, but prompts are task-specific.
• No universal “best prompt” — effectiveness varies by model and task.
• Automatically optimized prompts often outperform manually crafted ones.
• Optimized prompts can be surprisingly unconventional or eccentric.

Prompt optimization

Positive thinking

You are an experienced emergency room nurse. Take a deep breath and carefully assess the following patient’s symptoms.

Chain of Thought

Think through this patient’s case step-by-step: 1) List the symptoms, 2) Consider possible causes, 3) Evaluate urgency, 4) Recommend action.

Optimized prompt

The ER is in chaos, Doctor. We need your expertise to navigate this storm of patients and identify the most critical cases.

see Battle and Gollapudi (2024)

References

Battle, Rick, and Teja Gollapudi. 2024. “The Unreasonable Effectiveness of Eccentric Automatic Prompts.” February 20, 2024. https://doi.org/10.48550/arXiv.2402.10949.

Stechly, Kaya, Karthik Valmeekam, and Subbarao Kambhampati. 2024. “Chain of Thoughtlessness? An Analysis of CoT in Planning.” arXiv.org. May 8, 2024. https://arxiv.org/abs/2405.04776v2.