Blog

Prompt Engineering

Prompt Engineering, also known as In-Context Prompting, refers to methods for how to communicate with LLM to steer its behavior for desired outcomes without updating the model weights. It is an empirical science and the effect of prompt engineering methods can vary a lot among models, thus requiring heavy experimentation and heuristics.

This post only focuses on prompt engineering for autoregressive language models, so nothing with Cloze tests, image generation or multimodality models. At its core, the goal of prompt engineering is about alignment and model steerability.

The purpose of presenting few-shot examples in the prompt is to explain our intent to the model; in other words, describe the task instruction to the model in the form of demonstrations. However, few-shot can be expensive in terms of token usage and restricts the input length due to limited context length. So, why not just give the instruction directly?

Instructed LM (e.g. InstructGPT, natural instruction) finetunes a pretrained model with high-quality tuples of (task instruction, input, ground truth output) to make LM better understand user intention and follow instruction. RLHF (Reinforcement Learning from Human Feedback) is a common method to do so. The benefit of instruction following style fine-tuning improves the model to be more aligned with human intention and greatly reduces the cost of communication.

When interacting with instruction models, we should describe the task requirement in details, trying to be specific and precise and avoiding say “not do something” but rather specify what to do.

Self-Consistency Sampling

Self-consistency sampling (Wang et al. 2022a) is to sample multiple outputs with temperature > 0 and then selecting the best one out of these candidates. The criteria for selecting the best candidate can vary from task to task. A general solution is to pick majority vote. For tasks that are easy to validate such as a programming question with unit tests, we can simply run through the interpreter and verify the correctness with unit tests.

Chain-of-Thought (CoT)

Chain-of-thought (CoT) prompting (Wei et al. 2022) generates a sequence of short sentences to describe reasoning logics step by step, known as reasoning chains or rationales, to eventually lead to the final answer. The benefit of CoT is more pronounced for complicated reasoning tasks, while using large models (e.g. with more than 50B parameters). Simple tasks only benefit slightly from CoT prompting.