Don't Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration

Despite efforts to expand the knowledge of large language models (LLMs), knowledge gaps—missing or outdated information in LLMs—might always persist given the evolving nature of knowledge. In this work, we study approaches to identify LLM knowledge gaps and abstain from answering questions when knowledge gaps are present. We first adapt existing approaches to model calibration or adaptation through fine-tuning/prompting and analyze their ability to abstain from generating low-confidence outputs. Motivated by their failures in self-reflection and over-reliance on heldout sets, we propose two novel approaches that are based on model collaboration, i.e., LLMs probing other LLMs for knowledge gaps, either cooperatively or competitively. Extensive experiments with three LLMs on four QA tasks featuring diverse knowledge domains demonstrate that both cooperative and competitive approaches to unveiling LLM knowledge gaps achieve up to 19.3% improvements on abstain accuracy against the strongest baseline. Further analysis reveals that our abstention methods pinpoint failure cases in retrieval augmentation and knowledge gaps in multi-hop reasoning.1

Introduction. LLMs demonstrate impressive capabilities of encoding real-world knowledge in model parameters and leveraging it to aid knowledge-intensive tasks (Petroni et al., 2019; Brown et al., 2020; Yu et al., 2023a). But when such knowledge is missing or unreliable, they resort to hallucinations (Ji et al., 2023) and biases (Feng et al., 2023a), while still “speaking with confidence.” A growing body of work seeks to expand LLM knowledge through retrieval augmentation (Guu et al., 2020; Borgeaud et al., 2022; Khattab et al., 2022; Shi et al., 2023; Chen et al., 2023), search engine integration (Nakano et al., 2021; Press et al., 2023), and multi-LM collaboration (Luo et al., 2023; Feng et al., 2023b). However, LLM knowledge gaps might always persist due to the ever-evolving nature of knowledge (Kandpal et al., 2023; Mallen et al., 2023; De Cao et al., 2021; Hernandez et al., 2023; Kasai et al., 2024). Consequently, we posit that abstaining from generating low-confidence outputs should be a part of LLMs’ functionality, and ask a crucial research question: how to identify knowledge gaps in LLMs?

Discussion / Conclusion. We investigate AbstainQA, a setting where LLMs should abstain from answering questions incorrectly. We curate a taxonomy of 11 abstain baselines across four categories and propose COOP- ERATE and COMPETE, two novel abstain mechanisms that promote mechanistic reflection through multi-LLM collaboration, in cooperation or competition. Extensive experiments on four datasets demonstrate that COOPERATE and COMPETE advances the state-of-the-art in AbstainQA, with the potential to improve retrieval-augmented LLMs, multi-hop knowledge reasoning, and more.