1 of 19

Synchronous Faithfulness Monitoring for Trustworthy Retrieval-Augmented Generation

Di Wu, Jia-Chen Gu, Fan Yin, Nanyun Peng, Kai-Wei Chang

University of California Los Angeles

2 of 19

Motivation

  • Retrieval augmentation facilitates large language models (LLMs) in solving knowledge-intensive tasks.

​

Query

Up-to-date knowledge

LLMs

Private knowledge

Tools

Accurate and Personalized Response

3 of 19

Motivation

  • However, recent research find that retrieval-augmented language models (RALMs) are not faithful to knowledge.

[1] RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models

4 of 19

Research Questions

We investigate two core research questions:

​

​

Can we detect faithfulness issues from RALMs synchronously?

​

​

Can we synchronously improve the faithfulness of RALM decoding?

5 of 19

Proposal

6 of 19

Proposal

7 of 19

Proposal

8 of 19

Proposal

9 of 19

Proposal

10 of 19

Proposal

11 of 19

SynCheck

  • We propose to monitor synchronous features that can indicate four types of unfaithfulness behaviors.
  • Using a small amount of task-specific data, we train a lightweight MLP to aggregate these features at segment-level.

Behavior

Feature

Unknown knowledge

Likelihood

Unconfident use of knowledge

Local Intrinsic Dimension

Overdominance of parametric knowledge

Contrastive Context Influence

Misinterpretation of context

Semantic Alignment

12 of 19

Faithfulness-Oriented Decoding

  • How to improve the context faithfulness of RALM decoding?

​

  • Existing works
    • Abstention [2] is conservative and harms informativeness.
    • Reranking [3] lacks fine-grained control.
    • Single-feature contrastive decoding [4]

​

  • Can we do better with SynCheck?

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

[3] Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

[4] Trusting Your Evidence: Hallucinate Less with Context-aware Decoding.

13 of 19

Faithfulness-Oriented Decoding

  • We introduce faithfulness-oriented decoding (FOD):
    • Backtracking at unfaithful segments
    • Forward-looking beam search guided by SynCheck

​

14 of 19

Experimental Setup

  • We compile six long-form generation datasets in four tasks
    • Biography Generation
    • Open-domain QA
    • Summarization
    • Data-to-text

​

  • Segment-level faithfulness labels are generated by
    • Mapping human-annotated errors from RAGTruth [1]
    • An NLI model to check the outputs against retrieved contexts.

​

[1] RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models

15 of 19

Results

  • SynCheck significantly outperforms previous faithfulness checking baselines in terms of AUROC.

16 of 19

Results

  • Strong transferability of SynCheck train/test tasks.
    • Data-to-text is a strong source (train) task.
    • QA and biography are easier target tasks to transfer to.

​

17 of 19

Results

  • FOD achieves strong faithfulness results while maintaining good informativeness from its outputs.

18 of 19

Results

  • FOD outperforms the baseline methods at all output lengths.
  • Faithfulness@L = faithfulness of the first L sentences in the output.

19 of 19

Summary

Paper

Code

  • We systematically study detecting and correcting the faithfulness issues in RALM decoding.
  • The proposed SynCheck checker achieve state-of-the-art faithfulness detection results with only synchronous signals from RALMs.
  • The proposed FOD algorithm achieves strong faithfulness while maintaining the informativeness of the output.