27. Prompting Techniques

Learning objectives

• Understand what is prompt engineering and the most common techniques

Prompting

From Prompt Engineering by Lilian Weng

Prompt Engineering

Also known as In-Context Prompting

How to communicate with LLM to steer its behavior for desired outcomes without updating the model weights.

It is an empirical task (science?) and varies among models.

Techniques

• Zero-shot: Feed the task text to the model and ask for results.

• Few-shot: Input/output of high quality examples on the target task.

  • More token consumption.
  • Choice of prompt format, training examples, and order leads to different performance.
  • The selection of examples should be diverse and relevant.
  • Be careful of majority label bias and recency bias

• Instruction prompting: Finetune technique on a pretrained model.
  • Instructed LM Training with tuples of (task instruction, input, ground truth output)
  • RLHF (Reinforcement Learning from Human Feedback) is a common method.
  • It improves the model to be more aligned with human intention.

Instruction prompting

• Self-Consistency Sampling: Sample multiple outputs with temperature > 0 and select the best.
  • The criteria for selecting the best candidate can vary from task to task. General solution: majority vote.
• Chain-of-Thought (CoT): Generates a sequence describing reasoning logics step by step.
  • Steps also known as reasoning chains or rationales.
  • For complicated reasoning tasks.
    • Few-shot CoT: give examples
    • Zero shot: Use statements like Let's think step by step

Other techniques

• Self-Ask: Prompt the model to ask following-up to build the thought process iteratively.
• IRCoT (Interleaving Retrieval CoT) and ReAct (Reason + Act) combine iterative CoT prompting with queries to Wikipedia APIs to add to the context.

Self Ask

Automatic Prompt Design

Input: prompt (sequence of prefix tokens).

Optimization task: probability of getting a desired output given input.

How? Prefix tokens are optimized on the embedding space via gradient descent

Multiple methods: AutoPrompt (Shin et al., 2020), Prefix-Tuning (Li & Liang (2021), P-tuning (Liu et al. 2021), Prompt-Tuning (Lester et al. 2021).

APE (Automatic Prompt Engineer; Zhou et al. 2022): It uses a score function to select the best candidate over a pool of model-generated instruction candidates.

Other examples

• PAL (Program-aided language models; Gao et al. 2022) and PoT (Program of Thoughts prompting; Chen et al. 2022): ask LLM to generate programming language statements.

PoT

• TALM (Tool Augmented Language Models; Parisi et al. 2022): model augmented with text-to-text API calls. It builds API requests and calls them to append the result.