Enabling Large Language Models to Generate Text with Citations

Recent years have seen rapid progress in large language models (LLMs) such as GPT-4, LLaMA, and their instruction-tuned variants like Alpaca and Vicuna. These models excel in generating coherent, contextually relevant text, enabling applications in question answering, summarization, and dialogue systems. Early efforts like Retrieval-Augmented Generation (RAG) integrated external retrieval modules with generative models, significantly improving factual correctness. However, the evaluation of such systems remains predominantly manual, relying on human judgment or proprietary metrics, which hampers reproducibility and large-scale benchmarking. As models grow larger and more capable, ensuring their outputs are factually supported and citeable becomes increasingly critical, especially in high-stakes domains like healthcare, law, and scientific research. Existing evaluation methods lack comprehensive automation, particularly for long-form, multi-source responses, creating a bottleneck for systematic development and comparison.

Despite impressive capabilities, current LLMs struggle with generating responses that are both factually accurate and well-supported by citations. The main issues include poor retrieval quality, limited context windows restricting multi-source integration, and inadequate automatic evaluation tools. These bottlenecks hinder the development of trustworthy models, as manual evaluation is costly, inconsistent, and non-reproducible. Moreover, existing metrics often fail to capture the nuanced support that citations provide, leading to overestimation of model reliability. Addressing these challenges requires designing end-to-end systems that can effectively retrieve, synthesize, and verify information, along with automated metrics that align closely with human judgment.

This paper introduces ALCE, a comprehensive benchmark for automatic evaluation of LLMs’ ability to generate long, citation-supported answers. Key innovations include: 1) a multi-source dataset collection covering diverse question types and data sources, 2) a multi-dimensional evaluation framework combining fluency (via MAUVE), correctness (task-specific metrics), and citation support (via NLI), 3) the integration of NLI models like TRUE for automatic citation validation, 4) novel prompting strategies such as summarization, snippet extraction, and interactive retrieval to optimize information synthesis. These innovations collectively enable scalable, reproducible, and nuanced assessment of model performance in factual correctness and citation fidelity, addressing the limitations of prior single-metric or manual approaches.

• �� Data collection: Assemble datasets (ASQA, QAMPARI, ELI5) with diverse question types and sources, segment into passages.
• �� Retrieval: Use dense retrievers (GTR, DPR) and BM25 to fetch top-100 passages per question.
• �� Prompt design: Implement vanilla prompts, summaries, snippets, and interactive strategies to guide models.
• �� Generation: Use models like GPT-4, ChatGPT, LLaMA, with prompts instructing citation inclusion.
• �� Validation: Apply NLI models (TRUE) to verify if generated statements are supported by cited passages.
• �� Metrics: Measure fluency with MAUVE, correctness with exact match and claim recall, citation support with recall and precision.
• �� Optimization: Use reranking, passage summarization, and interactive checks to improve outputs.
• �� Evaluation: Compare models across datasets, analyze trade-offs, validate automatic metrics against human judgments.

Experiments involve testing multiple models (GPT-4, ChatGPT, LLaMA variants) on three datasets, measuring fluency, correctness, and citation quality. Different prompting strategies are compared, including vanilla, summarization, snippet, reranking, and interactive retrieval. Metrics are computed automatically and validated via human evaluation. Results show that models like GPT-4 achieve about 50% full citation support on ELI5, with strategies like reranking improving citation relevance. The experiments highlight the importance of retrieval quality and context size, revealing current limitations and guiding future improvements.

Models demonstrate high fluency but limited factual support, with only 50% of responses fully supported by citations on ELI5. Summarization and reranking improve correctness and citation relevance, respectively. Interactive retrieval strategies do not significantly outperform simpler prompts, indicating room for better retrieval and synthesis methods. The automatic metrics correlate strongly with human judgments, validating their use for large-scale benchmarking. These findings emphasize the need for enhanced retrieval systems and longer context models to improve factual accuracy and citation support.

The benchmark enables development of trustworthy AI systems in critical domains like medical diagnosis, legal advice, and scientific research, where factual correctness and traceability are vital. It also supports industry efforts to automate content verification, improve knowledge bases, and enhance user trust in AI-generated information. Long-term, the framework can guide the creation of models capable of multi-source reasoning, transparent citation, and factual auditing, transforming AI into a reliable knowledge partner.

Current retrieval quality limits factual support; models' limited context windows hinder multi-source synthesis; automatic metrics, despite validation, may not fully capture complex reasoning. These issues necessitate advancements in retrieval algorithms, model architectures, and validation techniques to realize fully trustworthy, scalable AI systems.