Knowledge Verification to Nip Hallucination in the Bud

TL;DR

The superior efficacy of KCA is demonstrated in reducing hallucinations across six benchmarks, utilizing foundation LLMs of varying backbones and scales, which confirms the effectiveness of mitigating hallucinations by reducing knowledge inconsistency.

Abstract

While large language models (LLMs) have demonstrated exceptional performance across various tasks following human alignment, they may still generate responses that sound plausible but contradict factual knowledge, a phenomenon known as hallucination. In this paper, we demonstrate the feasibility of mitigating hallucinations by verifying and minimizing the inconsistency between external knowledge present in the alignment data and the intrinsic knowledge embedded within foundation LLMs. Specifically, we propose a novel approach called Knowledge Consistent Alignment (KCA), which employs a well-aligned LLM to automatically formulate assessments based on external knowledge to evaluate the knowledge boundaries of foundation LLMs. To address knowledge inconsistencies in the alignment data, KCA implements several specific strategies to deal with these data instances. We demonstrate the superior efficacy of KCA in reducing hallucinations across six benchmarks, utilizing foundation LLMs of varying backbones and scales. This confirms the effectiveness of mitigating hallucinations by reducing knowledge inconsistency. Our code, model weights, and data are openly accessible at \url{https://github.com/fanqiwan/KCA}.

Figures (11)

• Figure 1: Illustration of the knowledge inconsistency phenomenon where alignment data contains knowledge not seen by foundation LLMs during pretraining, as exemplified by the recently introduced "Direct Preference Optimization" (in red) technique for LLM alignment.
• Figure 2: Hallucination rate (y-axis) of instruction-tuned LLMs of 7B, including Pythia, Llama-2, and Mistral, with different knowledge inconsistency percentages (x-axis) detected using KCA on various benchmarks.
• Figure 3: The overview of the proposed KCA approach to mitigate hallucinations through knowledge consistent alignment. KCA first detects knowledge inconsistency through formulated examinations (Left), followed by calibrating inconsistent alignment instances using open-book, discard, or refusal tuning (Right).
• Figure 4: The percentage (%) of the consistent subset $\mathcal{D}_{co}$ and the inconsistent subset $\mathcal{D}_{inc}$ out of the whole dataset $\mathcal{D}$ across various foundation LLMs and datasets.
• Figure 5: The average hallucination rate (%) of the instructions with non-refusal/refusal responses divided by refusal tuning. To measure the hallucination rate of the instructions with refusal responses, we employ the standard tuning baseline to generate the responses across different foundation LLMs and benchmarks. Results show that instructions with refusal responses exhibit considerably higher hallucination rates compared to those with non-refusal responses.