#DeepLearning #Prompting #In-ContextLearning

Prompting and In Context Learning

Issues with Fine-Tuning

• Need large task-specific datasets for fine-tuning 需要大量专业的数据集
• Train endlessly 不通用
  • Collect data for task $T_1$ , fine-tune model to solve task $T_1$
  • Collect data for task $T_2$ , fine-tune model to solve task $T_2$
  • ...
• Prone to overfitting 容易过拟合
  • Large models adapt to very narrow task distribution, which may exploit spurious correlations
• Finetuning large models is expensive to time, memory, and cost

[!question]+ How to adapt a pre-trained model without fine-tuning

Prompting

+ What is Prompting?

• Prompt-based learning (inference)
• A new paradigm in Deep Learning / Machine Learning (NLP, CV)
• Encouraging a pre-trained model to make particular predictions by providing a “prompt” that instructs the model to perform the task effectively

The General workflow of Prompting:

1. Prompt Addition
2. Answer Prediction (Search)
3. Post-process the answer

Prompt Addition

• Given input $x$, this component modifies the input into a prompt $x^{\prime }$

  • Append a textual string to the input

• Contains two step

  • Define a template with two slots
    • An input slot $[X]$ for input $x$
    • An answer slot $[Z]$ for an intermediate generated answer $z$ that will later be mapped into label
  • Fill slot $[X]$ with the input $x$

Answer Prediction

Post-process the answer

Design Considerations for Prompting

• Pre-trained model choice
• Prompt engineering
• Answer engineering
• Multi-Prompt learning

Pre-trained model choice

Prompt engineering

Answer engineering

Multi-Prompt learning

The Elements of Prompting

A prompt contains any of the following elements

• Instruction − The description of a specific task that you want the model to perform
• Context − External information or additional context that can steer the model to better responses
• Input data − The input or question that we are interested to find a response for
• Output Indicator − The type or format of the output

Applications of Prompting

• Text classification
• Text summarization
• Information extraction
• Question Answering
• Conversation
• Code generation
• Reasoning

Techniques of Prompting

• Zero-shot Prompting 无样本
• Few-shot Prompting 少样本
• Chain-of-Thought Prompting
• Self-Consistency
• Tree of Thoughts
• Multimodal CoT Prompting
• Active-Prompt
• Generate Knowledge Prompting
• Retrieval Augmented Generation
• Automatic Reasoning and Tool-use