Large Language Model Reference Augmentation Through Entailment Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) suffer from hallucinations and lack transparency in their output generation, leading to inaccurate and misleading information, with limited visibility into their operation and source material references.
Innovation Solution
A method is introduced to train LLMs by segmenting source and response content, identifying entailment pairs, and annotating training data with references, enabling the model to provide transparent and trustworthy outputs by indicating source material segments used in generating responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If LLMs are trained on vast amounts of textual data using self-supervised and semi-supervised methods, then the model's language understanding and generation capabilities are improved, but hallucinations and lack of transparency in output generation occur
Solution Approach 1:
The patent segments source content into distinct segments and matches them with corresponding response segments using entailment pairs. This segmentation approach allows the model to trace specific output portions to specific source material segments, providing transparency and reducing hallucinations while maintaining language generation capabilities.
Solution Approach 2:
The patent introduces an intermediary annotation layer that connects source content segments with response segments through entailment relationships. This intermediary structure enables verification of output validity by providing a traceable mapping between input source material and generated responses, thereby improving reliability without sacrificing adaptability.
2Productivity
If LLMs generate responses without providing source material references, then the generation process is simpler and faster, but users cannot verify the validity of model conclusions
Solution Approach 1:
The patent performs preliminary segmentation of source content and pre-establishes entailment pairs between source segments and response segments during the training data preparation phase. This preliminary action enables the model to quickly retrieve and reference relevant source material during inference without significantly impacting generation speed, while simultaneously providing verification capability for users.
3Use of energy by moving object
If traditional LLM training methods are used, then computational resources are efficiently utilized, but the cost of ownership increases due to hallucinations and lack of transparency
Solution Approach 1:
The patent enables the LLM to automatically generate annotations that link response segments to source content segments through entailment pairs. This self-service capability allows the model to autonomously provide transparency and verification information without requiring additional external verification systems, maintaining computational efficiency while improving reliability and reducing ownership costs.
Data Source
AI summary
A method for automatically providing references to source materials corresponding to generative AI outputs is disclosed. A training data item including source content and a response corresponding to the source content is received. The source content is segmented into a plurality of source content segments, and the response is segmented into a plurality of target segments. At least one entailment pair that includes a target segment included in the plurality of target segments and a source content segment included in the plurality of source content segments is identified. The training data item is annotated using the at least one entailment pair. The annotated training data item is provided to a large language model. The large language model is trained using the annotated training data item.