Effective Large Language Model Adaptation for Improved Grounding and Citation Generation

TL;DR

This work addresses the hallucination problem in LLMs by grounding responses in retrieved passages and providing citations. It introduces AGREE, a learning-based framework that tunes pre-trained LLMs to self-ground and extend grounding at test time via an iterative retrieval loop. The method builds grounding data automatically from unlabeled queries using an NLI verifier, enabling supervised fine-tuning that preserves base-model behavior while improving citation quality. Across five datasets and two LLMs, AGREE achieves superior grounding and citation precision compared to prompting and post-hoc baselines, and its test-time adaptation further enhances both grounding and answer correctness, with demonstrated generalization to out-of-domain data and model-agnostic applicability.

Abstract

Large language models (LLMs) have achieved remarkable advancements in natural language understanding and generation. However, one major issue towards their widespread deployment in the real world is that they can generate "hallucinated" answers that are not factual. Towards this end, this paper focuses on improving LLMs by grounding their responses in retrieved passages and by providing citations. We propose a new framework, AGREE, Adaptation for GRounding EnhancEment, that improves the grounding from a holistic perspective. Our framework tunes LLMs to selfground the claims in their responses and provide accurate citations to retrieved documents. This tuning on top of the pre-trained LLMs requires well-grounded responses (with citations) for paired queries, for which we introduce a method that can automatically construct such data from unlabeled queries. The selfgrounding capability of tuned LLMs further grants them a test-time adaptation (TTA) capability that can actively retrieve passages to support the claims that have not been grounded, which iteratively improves the responses of LLMs. Across five datasets and two LLMs, our results show that the proposed tuningbased AGREE framework generates superior grounded responses with more accurate citations compared to prompting-based approaches and post-hoc citing-based approaches

Figures (8)

• Figure 1: Our framework, Agree, combines tuning (Section \ref{['sec:tuning']}) and test time adaptation (Section \ref{['sec:tta']}) for better attribution and citation generation.
• Figure 2: Illustration of the tuning process. We sample responses from the base model, use an NLI model to add citations to the sampled responses, and tune the base model with the best-grounded response. We also show a concrete example of tuning data on the right.
• Figure 3: Illustration of the test-time adaptation mechanism. The adapted LLM retrieves from the corpus based on self-generated citation information to refine its response in an iterative way.
• Figure 4: Output examples of the proposed Agree framework with text-bison-001 as the base model. TTA is able to improve the response by retrieving more relevant information to precisely support a statement (see top) or finding more passages to generate a more complete response (see bottom).
• Figure 5: Zero-shot prompt template for sampling initial responses from the base LLM.