Stop Sequence: Definition and Examples

A stop sequence is a control mechanism used during the inference of a language model. It consists of one or more character strings defined by the user in the API call parameters. When the model generates one of these sequences in its response, generation immediately stops, and the text produced up to that point is returned as the result.

This mechanism is essential for structuring LLM outputs. Without stop sequences, a model may continue generating text beyond what is useful: it could invent additional dialogue turns, add unrequested examples, or simply produce redundant content. Stop sequences act as a clear and predictable end signal, allowing precise and well-delimited responses.

In practice, stop sequences are particularly useful in scenarios where the output format is structured. For example, in a multi-turn chat system, you can use 'Human:' as a stop sequence to prevent the model from simulating the user's response. In a data extraction context, you can define a delimiter like '---' so that the model stops after producing a single information block.

Most language model APIs (OpenAI, Anthropic, Mistral, etc.) allow specifying multiple stop sequences simultaneously, typically up to four. The stop sequence itself is not included in the final response, which simplifies post-processing of the generated text.