Large language model explainability via counterfactuals

A system comprising:

• at least one computer processor;
• a field LLM executing on the at least one computer processor;
• an enterprise application executing on the at least one computer processor;
• a developer application executing on the at least one computer processor;
• a domain corpus in a data repository stored on a physical storage device;
• an embedding store in the data repository;
• and
• and a counterfactuals generator executing on the at least one computer processor;
• wherein the developer application is configured to:
• obtain a prompt to a field large language model (LLM) and a corresponding original response generated by the field LLM from the enterprise application,
• select a first phrase from a selected plurality of phrases from the prompt;
• obtain, from the counterfactuals generator, a syntactic phrase set for the first phrase from the domain corpus,
• obtain, from the counterfactuals generator, a semantic phrase set for the first phrase from the domain corpus,
• cause the counterfactuals generator to select a counterfactual set including phrases co-occurring in the syntactic phrase set and the semantic phrase set,
• assign a counterfactual score for each counterfactual of the counterfactual set,
• select a first counterfactual from a counterfactual subset of the counterfactual set having highest counterfactual scores,
• replace the first phrase in the prompt with the first counterfactual to obtain a modified prompt and process the modified prompt by the field LLM to obtain a modified response, and
• store the first phrase and at least one strongest counterfactual of the counterfactual subset as a first phrase pair in the data repository.

A method comprising:

• obtaining a prompt to a field large language model (LLM) and a corresponding original response generated by the field LLM;
• selecting a plurality of phrases from the prompt;
• selecting a first phrase from the selected plurality of phrases;
• obtaining a syntactic phrase set for the first phrase from a domain corpus by:
• obtaining a first syntactic embedding corresponding to the first phrase from a syntactic embedding model,
• searching an embedding store for syntactic embeddings corresponding to a first plurality of phrases in the domain corpus that are similar to the first syntactic embedding, and
• selecting the first plurality of phrases in the domain corpus as the syntactic phrase set;
• obtaining a semantic phrase set for the first phrase from the domain corpus by:
• obtaining a first semantic embedding corresponding to the first phrase from a semantic embedding model,
• searching the embedding store for semantic embeddings corresponding to a second plurality of phrases in the domain corpus that are distant from the first semantic embedding, and
• selecting the second plurality of phrases in the domain corpus as the semantic phrase set;
• selecting a counterfactual set including phrases co-occurring in the syntactic phrase set and the semantic phrase set;
• assigning a counterfactual score for each counterfactual of the counterfactual set as a syntactic embedding cosine distance of the each counterfactual with respect to the first phrase, divided with a semantic embedding cosine distance of the each counterfactual with respect to the first phrase;
• selecting a first counterfactual from a counterfactual subset of the counterfactual set having highest counterfactual scores;
• replacing the first phrase in the prompt with the first counterfactual to obtain a modified prompt and processing the modified prompt by the field LLM to obtain a modified response;
• and
• storing the first phrase and at least one strongest counterfactual of the counterfactual subset as a first phrase pair in a data repository.