All files

Edit

Let's personalize your
AI Assistant!

I plan to use DocGPT for

How creative should DocGPT be?

Chain of Thought Prompting

Chain of Thought (CoT) prompting is a technique that leverages the capabilities of Large Language Models (LLMs) to generate reasoning traces or 'chains of thought'. These chains aid in answering questions or solving tasks that involve complex reasoning, such as arithmetic or commonsense reasoning.

CoT prompting with demonstrations often requires creating varied and effective examples by hand, which can sometimes result in less-than-ideal solutions. To address this, Zhang and his team in 2022 suggested a method that removes the need for manual work. They used a "Let's think step by step" prompt with Large Language Models (LLMs) to automatically produce reasoning chains, one step at a time.

However, this automated process might still produce errors in the reasoning chains. To lessen the impact of these errors, having a diverse range of demonstrations is important. The team proposed Auto-CoT, a method that generates diverse reasoning chains for demonstrations by sampling different types of questions.

Auto-CoT has two main steps:

Question Clustering: This involves grouping similar questions from a dataset into a few categories or clusters.
Demonstration Sampling: Here, a representative question is chosen from each cluster and its reasoning chain is generated using Zero-Shot-CoT with simple guidelines.

These guidelines could involve things like the length of questions (for example, 60 tokens) and the number of steps in the reasoning process (for example, 5 steps). This encourages the model to create demonstrations that are simple and accurate.

ReAct Prompting

ReAct is a technique that combines "acting" and "reasoning", similar to how humans learn new tasks and make decisions. This approach is used to enhance the capabilities of Large Language Models (LLMs).

LLMs can generate reasoning traces to answer complex questions, a method known as Chain-of-thought (CoT) prompting. However, CoT has limitations, such as its inability to access external information or update its knowledge, leading to potential errors.

ReAct overcomes these issues by allowing LLMs to generate reasoning traces and actions for a task, and interact with external environments like Wikipedia. This provides the model with more information for reasoning and enables dynamic planning.

For example, in a question-answering task, a question is passed to the model. The model then generates a series of thoughts and actions, and observes the environment for additional information. This process helps the model retrieve information to support its reasoning, while the reasoning guides what information to retrieve next.

To illustrate how ReAct works, let's consider an example from a training set. The model is given a question and generates a series of thoughts, actions, and observations to answer the question. These steps include decomposing the question, extracting information, performing reasoning, guiding search formulation, and synthesizing the final answer.

Different types of tasks require different prompt setups. For tasks where reasoning is crucial, multiple thought-action-observation steps are used. For decision-making tasks with lots of action steps, thoughts are used less frequently.

Generate-Knowledge Prompting

Large Language Models (LLMs) are continually being enhanced, and one popular method involves incorporating knowledge or information to make the model's predictions more accurate.

A paper by Liu et al. in 2022 explored a similar idea: Could the model generate knowledge before making a prediction? This approach involves creating knowledge to be used as part of the prompt. This technique can be particularly useful for tasks that require commonsense reasoning.

For example, consider a simple prompt about golf:

Prompt:
"Part of golf is trying to get a higher point total than others. Yes or No?"

The model might incorrectly answer "Yes." This mistake shows the limitations of LLMs when performing tasks that require more knowledge about the world.

So, how can we improve this with knowledge generation? The idea is to let the model generate relevant knowledge before answering the question, which can help it make more accurate predictions. This method, called Generate-Knowledge Prompting, can be a powerful tool for improving the performance of LLMs.

Few-Shot Prompting

Large Language Models (LLMs) have impressive abilities to answer questions without prior examples, a capability known as zero-shot. However, for more complex tasks, LLMs can sometimes fall short.

This is where few-shot prompting comes in. In this technique, we provide the model with a few examples, or 'demonstrations', to help it understand the task better and improve its performance. These demonstrations condition the model for future examples where we want it to generate a response.

For instance, consider an example where the task is to use a new word correctly in a sentence. We provide the model with a demonstration, explaining what the word means and how it's used in a sentence. The model can then use this information to correctly use another new word in a sentence.

Interestingly, it was found that providing the model with just one example (1-shot) can be enough for it to learn the task. For more difficult tasks, we can provide more demonstrations (3-shot, 5-shot, 10-shot, etc.).

There are also some tips for creating effective demonstrations. For example, the label space and the input text distribution specified by the demonstrations are both important. The format used can also significantly impact performance. Even using random labels can be better than no labels at all. Selecting random labels from a true distribution of labels (instead of a uniform distribution) can also be beneficial.

Make the most out of DocGPT

Upgrade to DocGPT Pro for unlimited messages to GPT-4, unlimited usage of tools & much more. Try it 30 Days for free & cancel any time.