Examples

Quickstart

• MExGen Quick Start: MExGen is used to attribute an importance score to each sentence in the input document that is summarized by an LLM.
• CELL Quick Start - CELL is used to generate a modified input prompt for an LLM that will result in a response that is contrasted to the original response.
• Token Highlighter Jailbreak Inspector Quick Start - Token Highlighter is used to compute the importance of sentences in the input prompt to an LLM contributing to affirmative responses, with demonstrations to inspect jailbreak prompts.

MExGen

• MExGen for Question Answering - MExGen is used to explain an LLM's response to a question in terms of a document provided as context to the LLM. The explanation consists of sentence-level and mixed word- and sentence-level attributions.
• MExGen for Retrieval Augmented Generation - MExGen is used to explain an LLM's response to a question in retrieval-augmented generation (RAG). The explanation shows which retrieved documents and which parts thereof are most important to the LLM.
• MExGen for Summarization - MExGen is used to explain an LLM's summarization of a document using both sentence-level and mixed word-, sentence-level attributions.

CELL

• CELL for Natural Language Generation - CELL and mCELL (a myopic version of CELL that is more expensive) are used to generate a modified input prompt for LLMs that will result in a response that is contrasted to the original response. This is demonstrated with a small and a larger LLM.

Token Highlighter

• Token Highlighter Jailbreak Inspector - Token Highlighter is used to identify important segments in the input prompt (tokens/words/sentences) contributing to affirmative responses, with demonstrations to inspect jailbreak prompts