Glossary · Category

Additional Model Concepts & Prompting Techniques

8 plain-English definitions.

Chain-of-verification

A prompting technique where a model generates an initial answer, then generates and answers verification questions about its own claims to catch and correct errors.

In-context learning

The general phenomenon where a model adapts its behavior to a new task purely from information provided in the prompt, without any parameter updates.

Meta-prompting

Using an LLM to generate or optimize prompts for another (or the same) LLM automatically rather than hand-writing them.

Program-aided language model (PAL)

A prompting technique where the model writes and executes code to solve a reasoning problem (e.g., arithmetic) instead of computing the answer directly in natural language.

RETRO (Retrieval-Enhanced Transformer)

A model architecture that conditions generation on retrieved text chunks from a large external database via cross-attention, reducing the need to memorize facts in weights.

Role-based prompting

Structuring a conversation into system/user/assistant roles, the standard chat format most modern LLM APIs expect.

Skeleton-of-thought

A prompting/decoding technique that first generates a high-level outline (skeleton) of a response, then expands each point in parallel to reduce generation latency.

Toolformer

A research approach where a model is trained to decide for itself when and how to call external tools (e.g., calculators, search) by self-supervising on its own annotated usage.