Image by ai%20prompting&position=20&from_view=search&track=ais&uuid=a658aa34-aa02-4ce1-a502-c03d3396f395″>freepik
Large language models (LLMs), such as OpenAI's GPT and Mistral's Mixtral, increasingly play an important role in the development of ai-powered applications. The ability of these models to generate human-like results makes them perfect assistants for content creation, code debugging, and other time-consuming tasks.
However, a common challenge faced when working with LLM is the possibility of encountering factually incorrect information, popularly known as hallucinations. The reason for these events is not far-fetched. LLMs are trained to provide satisfactory responses to prompts; in cases where they cannot provide one, they invoke it. Hallucinations can also be influenced by the type of inputs and biases used in training these models.
In this article, we will explore three advanced, research-backed stimulation techniques that have emerged as promising approaches to reducing the occurrence of hallucinations while improving the efficiency and speed of results produced by LLMs.
To better understand the improvements that these advanced techniques bring, it is important that we talk about the basics of speed typing. Cues in the context of ai (and in this article, LLMs) refer to a group of characters, words, tokens, or a set of instructions that guide the ai model in terms of the human user's intent.
Cue engineering refers to the art of creating cues with the goal of better directing the behavior and resulting outcome of the LLM in question. By using different techniques to better convey human intent, developers can improve model results in terms of accuracy, relevance, and consistency.
Here are some essential tips to follow when crafting a message:
- Be concise
- Provide structure by specifying the desired output format.
- Please provide references or examples if possible.
All of this will help the model better understand what it needs and increase the chances of getting a satisfactory answer.
Below is a good example that queries an ai model with a message using all the tips mentioned above:
Prompt = “You are an expert ai prompting engineer. Please generate a two-sentence summary of the latest advances in prompt generation, focusing on the challenges of hallucinations and the potential of using advanced prompting techniques to address these challenges. The result must be in markdown format.”
However, following these essential tips discussed above does not always guarantee optimal results, especially when it comes to complex tasks.
Leading researchers from leading ai institutions such as Microsoft and Google have invested a lot of resources in optimizing LLM, that is, actively studying the common reasons for hallucinations and finding effective ways to address them. The following guidance techniques have been found to provide better and more context-aware instructions to the studied LLMs, increasing the chances of obtaining better relevant results and also reducing the probability of obtaining inaccurate or meaningless information.
Below are some examples of advanced, research-based stimulation techniques:
1. Incitement to emotional persuasion
TO 2023 study by Microsoft researchers found that using emotional language and persuasive prompts, called “EmotionPrompts,” can improve LLM performance by more than 10%.
This style adds a personal and emotional element to the given prompt, transforming the request into a very important one with significant consequences for the results. It's almost like talking to a human; Using an emotional angle helps communicate the importance of the task, stimulating deeper focus and commitment. This strategy can be useful for tasks that require greater creativity and problem-solving skills.
Let's look at a simple example where emotion is used to enhance the message:
Basic notice: “Write a Python script to sort a list of numbers.”
Indicate with emotional Persuasion: “Excited to improve my Python skills, I need to write a script to sort numbers. This is a crucial step in my career as a developer.”
While both prompt variations produced similar code results, the “EmotionPrompts” technique helped create cleaner code and provided additional explanations as part of the generated output.
Another interesting experiment Finxter found that providing monetary tips to LLMs can also improve their performance, almost like appealing to a human being's financial incentive.
2. Chain of Thought Prompts
Another incitement technique discovered to be effective by a group of researchers from the University of Pittsburgh It is the Chain of Thought style. This technique uses a step-by-step approach that guides the model through the desired output structure. This logical approach helps the model craft a more relevant and structured answer to a complex task or question.
Below is an example of how to create a Chain of Thought style message based on the provided template (using OpenAI's ChatGPT with GPT-4):
Basic notice: “Write a digital marketing plan for a financial application aimed at small business owners in big cities.”
north
Chain of thought message:
“Describes a digital marketing strategy for a financial app for small business owners in big cities. Focus on:
- Select digital platforms that are popular with this business demographic.
- Create engaging content such as webinars or other relevant tools.
- Generate profitable tactics unique to traditional ads.
- Tailor these tactics to the needs of urban small businesses in a way that increases customer conversion rates.
Name and detail each part of the plan with unique, actionable steps.”
The Prompt Chain technique generated a more accurate and actionable result from a superficial look.
The Step-Back-Prompting technique, presented by seven from Google Deep Mind Researchers, is designed to simulate reasoning when it comes to LLM. This is similar to teaching a student the underlying principles of a concept before solving a complex problem.
To apply this technique, it is necessary to point out the underlying principle behind a question before asking the model to provide an answer. This ensures that the model gets solid context, which will help it give a technically correct and relevant answer.
Let's examine two examples (using OpenAI's ChatGPT with GPT-4):
Example 1:
Basic notice: “How do vaccines work?”
Directions for using the step back technique
- “What biological mechanisms allow vaccines to protect against disease?”
- “Can it explain the body's immune response triggered by vaccination?”
While the basic question provided a satisfactory answer, the use of the step-back technique provided a more in-depth and technical answer. This will be especially helpful for technical questions you may have.
As developers continue to create novel applications for existing ai models, there is a growing need for advanced cueing techniques that can enhance the capabilities of large language models to understand not only our words but also intent and emotion behind them to generate more precise and contextually relevant words. Departures.
