LLM Attribution Verification via Quote Source Checking

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Large language models (LLMs) often generate hallucinated content when summarizing documents, leading to inaccurate and unreliable results, as they may produce information not supported by the source document.

Innovation Solution

The implementation of programmatic solutions that generate specific prompts for LLMs to produce verbatim quotes and include source indicators, along with post-processing verification to check these quotes against the source document, reducing the likelihood of hallucinations and ensuring accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If LLM is used to generate summaries and answers from source documents, then productivity is improved by reducing time to read and comprehend documents, but reliability deteriorates due to hallucinated content that is not supported by the source document

Engineering Contradiction:

Improvetime to read and comprehend documentsVSAvoidaccuracy of summary content

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system implements feedback by verifying LLM output against the source document. A verification process checks whether generated content is actually supported by the source document, and visual indicia are provided to indicate verification status. This closed-loop feedback mechanism resolves the contradiction by maintaining productivity while improving reliability through continuous verification.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system applies preliminary action by generating specific prompts that instruct the LLM to produce only verbatim quotes from the source document. This preemptive measure reduces hallucinations before they occur, allowing the system to maintain both high productivity and reliability by preventing inaccurate content generation in the first place.

Inventive Principle:

Principle #10Preliminary action

2Reliability

If prompts instruct LLM to produce only verbatim quotes, then reliability is improved by reducing hallucinations, but device complexity increases due to additional verification steps

Engineering Contradiction:

Improveaccuracy of LLM outputVSAvoidcomplexity of verification system

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The system applies self-service by having the LLM include source indicators and verbatim quotes that directly reference the source document. The verification process leverages these self-provided references to check accuracy, reducing the need for complex external verification mechanisms while maintaining high reliability.

Inventive Principle:

Principle #25Self-service

3Reliability

If visual indicia and links are added to verify quotes, then reliability is improved by indicating verification status, but ease of operation deteriorates due to additional information to process

Engineering Contradiction:

Improvetrustworthiness of outputVSAvoidsimplicity of output display

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

The system applies local quality by providing visual indicia and links only at specific locations where verification is relevant - adjacent to individual quotes and statements in the LLM output. This localized approach maintains reliability through verification indicators while preserving ease of operation by avoiding unnecessary information and keeping the interface clean and focused.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS20240296295A1Attribution verification for answers and summaries generated from large language models (LLMS)

Publication Date: 2024.09.05 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

Systems and methods for verifying attribution of quotations, generated by a large language model (LLM), to a source document are disclosed herein. Upon a request to summarize a source document or process a question that is answerable from a document, an LLM prompt is formed with the request or question along with the content of the source document. The LLM prompt is configured to cause an LLM to generate quotes that are intended to be from the source document. The output of the LLM, including the quotes, is then verified against the source document.