With the ChatGPT craze sparked by OpenAI, there is a newly invented job called prompt engineer.
This is a new role that handles the prompt-engineering process of guiding generative AI to produce the results you want. The job is designing more detailed instructions so that generative AI can produce high-quality, highly relevant output.
Perhaps because of this, bookstores now carry many prompt engineering and ChatGPT-related books, and the number keeps growing.
Prompt engineering is the work of designing and optimizing prompts that guide a generative AI model to perform a specific task. Doing so can improve the model performance and accuracy. There are various prompt techniques; below are a few.
1. Instruction: clearly describe the specific task or command the language model must execute.
2. Context: provide external information or additional context so the language model can precisely understand and respond to the command.
3. Input data: provide the information on which the model will base its output. It may include the content or question provided by the user.
4. Output specification: specify the desired output type or format.
Source: conversation with Bing
By combining these elements, you can design a prompt and guide the model behavior. So I asked Copilot a bit more about various prompt techniques.
- Zero-shot Prompting: Describe the content and expected result without providing a specific example. Large models can handle zero-shot prompting effectively, but sometimes results fall short of expectations.
- Few-shot Prompting: Provide a few examples so the large model can generate more accurate results. Useful especially for problems such as arithmetic and logical reasoning.
- Chain of Thoughts Prompting: Improves complex reasoning by introducing intermediate reasoning steps. It involves explaining a series of intermediate steps and manually composing the reasoning process.
- Active-Prompt: An advanced technique of designing prompts that guide the model to produce the desired output. It helps the model generate the target result more accurately. Works together with Chain-of-Thought (CoT) to activate intermediate reasoning steps.
- Directional Stimulus Prompting: A technique of designing prompts that steer the model in a specific direction. For example, when asking about a particular topic, you can guide the model to focus on that topic.
- ReAct: A prompting approach for adapting large language models to various tasks. It tunes the model using human-designed CoT reasoning together with task-specific example prompts.
- Multimodal CoT: A method of designing prompts that combine multiple modalities (e.g., text, images). For example, a law student may structure a reasoning process using both images and text.
Prompt engineering helps researchers raise the capability of generative AI models and adapt them to various tasks and application domains. In conclusion, prompt engineering lets you get the expected results without updating the model itself.
Right.
..
It used to.
...
It used to, that is.. ?
?!
Recently, because of a paper released by Google, I cautiously predict that the prompt engineer(ing) field will be re-evaluated. The paper is a study of SELF-DISCOVER - having the language model itself discover a reasoning structure for solving complex problems.
By discovering its own reasoning structure, the language model goes beyond the classic direct-answer approach and prompting methods like Chain-of-Thought (CoT), and automatically composes a multi-step reasoning process. Through this, tailored to each model characteristic, the model performs self-learning to interpret and reason about the problem itself and repackage it into an optimal prompt - without the user having to do prompt engineering by hand.
In short, you can say it sloppily and it still answers on point. ;)
Self-Discover: Large Language Models Self-Compose Reasoning Structures
We introduce SELF-DISCOVER, a general framework in which LLMs self-discover task-intrinsic reasoning structures to tackle complex reasoning problems that are difficult for typical prompting methods. At the core is a self-discovery process in which the LLM selects several atomic reasoning modules such as critical thinking and step-by-step thinking, and composes them into an explicit reasoning structure that the LLM then follows during decoding. SELF-DISCOVER greatly improves GPT-4 and PaLM 2 on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH - by up to 32% compared with Chain of Thought (CoT). It also outperforms reasoning-heavy methods such as CoT-Self-Consistency by more than 20% while requiring 10 to 40 times less reasoning compute. Finally, we show that self-discovered reasoning structures are universally applicable across model families - from PaLM 2-L to GPT-4 to Llama2 - and share commonalities with human reasoning patterns.
A diagram of using self-discovery for problem solving. Given a generator LM, a task, and descriptions of seed reasoning modules, we guide the LM to generate a reasoning structure in key-value form to solve the task. Finally, the model can solve every instance of the task by following the self-discovered structure and filling in values in JSON step by step.
Comparison of reasoning processes generated by CoT, Plan-and-Solve, and Self-Discover on a BBH geometric-shapes task. Both CoT and Plan-and-Solve incorrectly judge that the path is not closed and therefore does not form a regular shape, arriving at the wrong answer. Self-Discover reasoning structure first breaks down each line segment, analyzes the coordinates carefully, and then uses logical reasoning to conclude that the path ends at the same coordinates and therefore forms a closed shape, finally reaching the correct answer through final reasoning.
Source: original paper
https://arxiv.org/abs/2402.03620
