Presented at ACL 2024, the AGREE (Adaptation for GRounding EnhancEment) framework represents a collaborative effort between UT Austin and Google Cloud AI. AGREE addresses the critical need for reliability in Large Language Models (LLMs) by training them to self-ground claims and provide accurate, verifiable citations.
The Grounding Challenge in RAG
Retrieval-Augmented Generation (RAG) is a common technique to mitigate hallucinations by providing external context. However, current RAG systems face two primary limitations:
- Citation Hallucination: Standard prompting often leads to citations that do not actually support the generated claims or point to incorrect passages.
- Static Retrieval: Most systems retrieve documents once. If the initial retrieval is insufficient, the model lacks a mechanism to seek further information, leading to failure or fabrication.
Existing post-hoc solutions attempt to verify citations after generation, but these are computationally expensive and rely heavily on the model's internal knowledge during the initial generation phase.
The AGREE Framework
AGREE is a learning-based framework that integrates grounding directly into the model's generation process. It consists of two distinct phases: Training Adaptation and Test-Time Adaptation.
Training Adaptation: Learning to Cite
The training phase focuses on fine-tuning the LLM to generate grounded responses with inline citations. To avoid the high cost of human labeling, the authors developed an automated "self-improvement" loop:
- Data Construction: Candidate responses are generated from unlabeled queries using a base LLM.
- NLI Filtering: An external Natural Language Inference (NLI) model (specifically the TRUE model) evaluates every sentence against retrieved passages.
- Fine-Tuning: The model is fine-tuned using Parameter-Efficient Fine-Tuning (LoRA) on the highest-quality grounded data. A key innovation is teaching the model to explicitly list "Claims Not Supported" at the end of its response.
Test-Time Adaptation (TTA)
During inference, AGREE employs an iterative refinement strategy. If the model identifies unsupported statements in its initial output, it uses those statements as specific search queries to retrieve better evidence. The response is then regenerated, continuing for up to four iterations or until all claims are grounded.
Experimental Evaluation
The researchers evaluated AGREE across five diverse datasets, including fact-checking (FEVER), multi-hop reasoning (StrategyQA), and ambiguous questions (ASQA). Testing also included an out-of-distribution (OOD) proprietary Enterprise dataset for customer support.
Key Insights and Results
- Precision and Recall: AGREE achieved over a 20% improvement in citation precision and recall compared to standard prompting-based approaches.
- Self-Awareness as a Trigger: Training the model to recognize what it doesn't know (unsupported claims) is crucial for triggering the iterative TTA loop.
- Effective Distillation: Smaller models, such as Llama-2-13b, were successfully tuned using data generated by larger models like Google's
text-bison, allowing them to rival the performance of much larger proprietary models. - Zero-Shot Generalization: Models tuned on general Wikipedia-based tasks demonstrated strong grounding capabilities when applied to the specialized Enterprise dataset without further domain-specific tuning.
Conclusion
AGREE demonstrates that grounding and citation generation are learnable skills that significantly enhance LLM reliability. By combining automated data construction with iterative test-time refinement, the framework provides a scalable path toward more trustworthy AI systems that can verify their own claims.
Summary updated