-
Background Large Language Models (LLMs) have been revolutionary in natural language processing by offering a unified architecture for diverse tasks. However, they are prone to generate hallucinated content that can mislead users.
-
Existing Work Existing methods for hallucination detection in LLMs are limited due to a lack of a fine-grained checking mechanism and a comprehensive benchmark for measuring hallucinations across different real-world applications.
- Introducing REFCHECKER framework
- Challenge 1: Fine-grained hallucination detection Existing granularity for detecting hallucinations in responses is insufficient, especially for long and complex ones. The researchers propose using "claim-triplets" as the units of detection to achieve finely grained semantic partitioning and detection, significantly improving over sentence-level and sub-sentence level granularity.
- Challenge 2: Comprehensive benchmark There is a lack of comprehensive benchmarks to assess hallucinations across multi-tasking. The REFCHECKER benchmark includes three classes of real-world LLM tasks and has manually annotated 11,000 claim-triplets from the responses of 7 LLMs, ensuring high-quality evaluation data.
The REFCHECKER framework consists of two main components: an extractor that generates claim-triplets and a checker that evaluates each claim-triplet against a reference. It is highly automated and scalable for hallucination detection across various tasks. REFCHECKER has shown an improvement of 6.8 to 26.1 points over prior methods in evaluations and supports both proprietary and open-source models.
REFCHECKER is a framework that detects fine-grained hallucinations in LLMs and benchmarks them. It detects and verifies factual inconsistencies in responses with high precision and strong alignment with human judgments.