LLM Self-Grounding via Iterative Attribution Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) often hallucinate by generating plausible but non-factual information, limiting their applicability in real-world settings, and existing solutions for mitigating hallucinations are either unsatisfactory or not extendable to specific and narrow knowledge fields.
Innovation Solution
The system adapts an LLM to self-ground its responses by attributing them to passages and evaluating their reliability, using an iterative inference process to refine results, thereby improving grounding quality and reducing hallucinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If LLMs generate responses without grounding, then generation speed and fluency are improved, but hallucination increases and reliability decreases
Solution Approach 1:
The patent merges the LLM generation capability with grounding verification capability into a unified system. The LLM both generates responses and verifies their grounding against retrieved passages, eliminating the need for separate post-hoc citation models and reducing overall system complexity while improving reliability.
Solution Approach 2:
The LLM performs self-verification of its own responses by checking whether generated statements are supported by retrieved passages. This self-grounding mechanism allows the model to autonomously ensure response reliability without external intervention or complex post-processing systems.
2Reliability
If post-hoc citing is used to link support passages to claims, then grounding quality improves, but the system becomes heavily reliant on parametric knowledge and fails to extend to specific knowledge fields
Solution Approach 1:
The system performs preliminary retrieval of relevant passages from an external knowledge base before the LLM generates its response. This preliminary action ensures that the LLM has access to accurate, up-to-date information specific to the knowledge field being queried, enabling it to adapt to domain-specific requirements without relying solely on its pre-trained parametric knowledge.
Solution Approach 2:
The patent introduces an intermediary retrieval mechanism that bridges the LLM and the knowledge base. This intermediary component selectively retrieves relevant passages based on the query, allowing the system to adapt to specific knowledge fields by filtering and selecting only the most relevant information for the given task.
3Ease of manufacture
If instruction tuning or in-context learning is used to generate grounded responses, then grounding capability is introduced, but grounding quality remains unsatisfactory when following retrieval-augmented generation framework
Solution Approach 1:
The system implements a feedback mechanism where the LLM generates responses, then self-verifies them against retrieved passages, and uses this verification feedback to refine its output. This iterative feedback loop continuously improves grounding quality by allowing the model to learn from its own performance and adjust its responses accordingly.
Solution Approach 2:
The patent creates a composite system that combines multiple capabilities: LLM generation, retrieval augmentation, and self-verification. This composite approach integrates the strengths of each component while mitigating their individual weaknesses, achieving superior grounding quality compared to any single approach used in isolation.
4Reliability
If external models are invoked for citation generation, then grounding accuracy improves, but system overhead and computational cost increase
Solution Approach 1:
The LLM performs self-verification of its own responses by checking whether generated statements are supported by retrieved passages. This self-grounding mechanism allows the model to autonomously ensure response reliability without external intervention or complex post-processing systems.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for grounding LLMs. In one aspect, a method includes accessing responses from a large language model, each response comprising data including: a query, an answer to the query, the answer comprising one or more statements, citations linking each statement to an evidence passage in a corpus; determining a grounding quality of the answer based on the evidence passages and the statements using an attribution evaluation model, wherein the grounding quality is a quantification of attribution of the statements in the answer to a document corpus; and tuning the large language model to obtain an adapted large language model that satisfies a grounding constraint based on grounding quality score.