Highlights
We introduce SynCheck, an efficient synchronous checker for faithful retrieval-augmented langauge models (RALMs). We further introduce FOD, a faithfulness-oriented decoding mechanism that guides RALM decoding on-the-fly.
We introduce SynCheck, an efficient synchronous checker for faithful retrieval-augmented langauge models (RALMs). We further introduce FOD, a faithfulness-oriented decoding mechanism that guides RALM decoding on-the-fly.
Retrieval-augmented language models (RALMs) have shown strong performance and wide applicability in knowledge-intensive tasks. However, there are significant trustworthiness concerns as RALMs are prone to generating unfaithful outputs, including baseless information or contradictions with the retrieved context.
Existing post-hoc approaches solve faithfulness issues via citation or revision, yet they are computationally costly. Meanwhile, existing on-the-fly interventions adopted by RALM systems are accuracy-driven, while their efficacy in detecting and correcting faithfulness errors is unknown. In this work, we ask:
We build SynCheck, an accurate checker for faithfulness errors that ensembles a range of efficiently measurable decoding dynamics. On a suite of six long-form retrieval-augmented generation datasets, SynCheck achieves 0.85 AUROC in detecting faithfulness errors across six long-form retrieval-augmented generation tasks, improving prior best method by 4%.
Although SynCheck requires training a lightweight aggregator, we find that the training processs does not need to be task-specific.
In addition, we can train SynCheck on a surrogate model, which can be leveraged to detect faitfhulness errors in other models' outputs.
We further introduce FOD, a faithfulness-oriented decoding algorithm that levereages SynCheck's on-the-fly feedback with a beam search algorithm.
On a suite of six datasets, we find FOD significantly improves over a range of baseline decoding strategies in improving the faithfulness of the output.
In addition, compared to CAD, the prior state-of-the-art faithful decoding method, FOD achieves a better faithfulness at the first L sentences, with L=1, ..., 10.
@misc{wu2024synchronous,
title={Synchronous Faithfulness Monitoring for Trustworthy Retrieval-Augmented Generation},
author={Di Wu and Jia-Chen Gu and Fan Yin and Nanyun Peng and Kai-Wei Chang},
year={2024},
eprint={2406.13692},
archivePrefix={arXiv},
primaryClass={id='cs.CL' full_name='Computation and Language' is_active=True alt_name='cmp-lg' in_archive='cs' is_general=False description='Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.'}
}