LLM Hallucination Reduction via Search-Mediated Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models often generate factually incorrect content, known as hallucinations, which can occur even when prompted with factual contexts, leading to unreliable responses.
Innovation Solution
A system that incorporates a query generator and a response generator, where the query generator identifies relevant supporting resources through search engines and fine-tunes the model to use these resources appropriately, reducing hallucinations by learning when and how to rely on evidence-based information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large language model generates creative content without external constraints, then creativity and flexibility are improved, but factual accuracy deteriorates due to hallucinations
Solution Approach 1:
The patent introduces search engines as an intermediary between the language model and factual information. The search engine retrieves external knowledge and provides it to the model, acting as a mediator that connects the creative generation process with factual accuracy without constraining the model's creativity
Solution Approach 2:
The system performs preliminary information retrieval through search engines before the language model generates its response. By obtaining relevant factual information in advance and providing it as context to the model, the system ensures factual accuracy is established before creative generation occurs
2Reliability
If search results are provided to the model as input, then factual accuracy is improved, but device complexity increases due to additional components
Solution Approach 1:
The language model serves multiple functions: it acts as both the creative content generator and the processor of search results. By making the model multi-functional, the system avoids adding separate dedicated components for processing external information, thereby limiting the increase in system complexity
3Reliability
If the model is fine-tuned to learn when to use search results, then hallucinations are reduced, but training time and computational resources increase
Solution Approach 1:
Instead of fine-tuning the model on all possible scenarios, the system applies partial action by only fine-tuning on specific datasets that demonstrate when to use search results and when to rely on internal knowledge. This selective approach reduces training time while still achieving the goal of reducing hallucinations
Data Source
AI summary
Implementations relate to helping a large language model generate factual responses to prompts that request factual content is disclosed. The large language model may receive a prompt context, a plurality of encoded context passages as input. The large language model is trained to determine whether or not to utilize the encoded context passages in generating the response. Implementations also relate to different methods of fine-tuning the responses generated by the large language model through query refinements, response re-writes, and evaluation of factual accuracy.