A user provides the following text to a large language model: 'Write a poem about the ocean.' This input is an example of a discrete, human-readable prompt. Explain the fundamental transformation this prompt undergoes inside the model before the generative process begins.

Google

Hard prompts, also known as discrete prompts, are explicit, predefined sequences of tokens, typically in natural language, that users input directly to guide a Large Language Model. These human-readable prompts serve as the standard input for LLMs, which internally encode the discrete sequences into low-dimensional, real-valued vectors before generating a response.

Discrete Prompts (Hard Prompts)

Generally, these methods scrapes a large data corpus for strings containing $x$ and $y$, and finds either the middle words or dependency paths between the inputs and outputs. Frequent middle words or dependency paths can serve as a template.

Prompt Mining

Treating prompt generation as a paraphrasing task is a method used during the expansion step of prompt optimization. It involves applying off-the-shelf paraphrasing systems, which can be based on Large Language Models or other architectures, to transform an input prompt into new, semantically equivalent candidate prompts to explore the search space.

Prompt Paraphrasing

These methods applies a gradient-based search over actual tokens to find short sequences that can trigger the underlying pre-trained LM to generate the desired target prediction.

Gradient-based search

These works treat the generation of prompts as a text generation task and use standard natural language generation models to perform this task.

 Prompt Generation

These works investigate the task of knowledge base completion and design a template for an input (head-relation-tail triple) using LMs.

Prompt Scoring

A key research question in automated prompting investigates whether more compact and efficient prompt representations exist beyond traditional discrete text strings. This inquiry is motivated by the fact that LLMs internally encode discrete prompts (or "hard prompts") into low-dimensional vectors, suggesting that a more direct, non-string representation could be possible and more efficient.

Exploring and Learning Non-String Prompt Representations

Continuous prompts, also known as soft prompts, are implicit and adaptable prompting patterns that exist as hidden, distributed representations directly within the embedding space of a Large Language Model. Unlike human-readable hard prompts, they are not required to correspond to meaningful text. Instead, soft prompts are treated as learnable parameters that are optimized through continuous optimization, enabling the exploration of prompting methods beyond discrete, user-defined text.

Continuous Prompts (Soft Prompts)

A hard prompt is represented by the variable $c$, and its corresponding sequence of vector embeddings is denoted as $e_c$.

Notation for Hard Prompts and Their Embeddings

A developer is trying to make a language model summarize articles. They provide the model with the following text input for each article: 'Summarize the following text in three sentences: [article text]'. Which of the following statements best analyzes why this input is considered a 'discrete' or 'hard' prompt?

Based on your understanding of how language models process user inputs, evaluate the team's proposed strategy. Identify one primary advantage and one significant disadvantage of relying solely on these explicit, predefined text inputs.

Evaluating a Prompting Strategy for a Customer Service Chatbot

Analyzing the Transformation of a User Prompt

An example of a hard prompt—an explicit, natural language text sequence input by a user to guide a model's response—is the instruction: "Translate the sentence into Chinese. Consider it done!" This illustrates how hard prompts are formulated in human-readable language to communicate tasks directly.

Learn Before

Related