Attribute or Abstain: Large Language Models as Long Document Assistants

Abstract: LLMs can help humans working with long documents, but are known to hallucinate. Attribution can increase trust in LLM responses: The LLM provides evidence that supports its response, which enhances verifiability. Existing approaches to attribution have only been evaluated in RAG settings, where the initial retrieval confounds LLM performance. This is crucially different from the long document setting, where retrieval is not needed, but could help. Thus, a long document specific evaluation of attribution is missing. To fill this gap, we present LAB, a benchmark of 6 diverse long document tasks with attribution, and experiments with different approaches to attribution on 5 LLMs of different sizes. We find that citation, i.e. response generation and evidence extraction in one step, performs best for large and fine-tuned models, while additional retrieval can help for small, prompted models. We investigate whether the "Lost in the Middle'' phenomenon exists for attribution, but do not find this. We also find that evidence quality can predict response quality on datasets with simple responses, but not so for complex responses, as models struggle with providing evidence for complex claims.

• The paper demonstrates that citation methods yield superior evidence quality in LLM responses for long documents, particularly with models like GPT-3.5 and GPT-4.
• It evaluates multiple attribution techniques, revealing that context reduction can disrupt document coherence and limit LLM performance.
• The study highlights a strong correlation between evidence quality and response accuracy, paving the way for more trustworthy AI applications in complex tasks.

Overview of "Attribute or Abstain: LLMs as Long Document Assistants"

"Attribute or Abstain: LLMs as Long Document Assistants" investigates how LLMs can efficiently assist with tasks involving long documents while maintaining high trust and verifiability through attribution. The study evaluates several attribution approaches on diverse LLMs and tasks, proposing LAB, a comprehensive benchmark for long document attribution. The work examines the efficacy of citation-based methods, investigates potential positional biases, and analyzes the correlation between the quality of evidence and responses.

Methods and Approaches

Attribution Techniques

The paper explores several attribution methodologies applicable to long document tasks:

1. Post-hoc: Generates a response first, then retrieves supporting evidence.
2. Retrieve-then-read: Selects evidence from the document prior to generating a response.
3. Citation: Executes response generation and evidence retrieval concurrently.
4. Reduced-variants: Employs initial retrieval to focus on a subset of the document, reducing input length.

Benchmark and Model Selection

LAB is designed to test the outlined approaches using six datasets covering QA, classification, NLI, and summarization, each with defined domains like science, law, and governmental documents. The study compares four LLMs: GPT-3.5, GPT-4, Flan-T5-XL, and Longchat, utilizing both prompted and fine-tuned models.

Figure 1: The approaches to attribution in long document scenarios analyzed in this work.

Key Findings and Insights

Optimal Attribution Approaches

• Citation methods generally yielded superior evidence quality compared to other approaches across most tasks, particularly when using large models like GPT-3.5 and GPT-4. This affirms the capability of LLMs to effectively manage retrieval tasks internally.
• Post-hoc methods showed improved performance in summarization tasks and when using smaller models like Longchat.
• The context reduction strategy did not consistently enhance performance, suggesting long documents' logical coherence might be disrupted, affecting LLM processing.

Positional Bias Analysis

• No significant "Lost in the Middle" phenomenon was observed; however, response quality often diminished when evidence appeared later in documents, indicating a positional performance dependency.
• The evidence distribution analysis reflected that models followed gold distribution patterns except for tasks like GovReport, where biases towards document beginnings were visible.

Figure 2: Evidence distribution and response quality analysis across document positions.

Evidence and Response Quality Correlation

• Strong correlation between evidence and response quality for simpler, single fact responses was identified, where evidence quality could predict response quality effectively.
• For datasets requiring multi-fact responses, models struggled to provide comprehensive evidence, leading to insufficient prediction accuracy compared to simpler tasks.

  Figure 3: Gold class distribution in selective prediction for GPT-3.5-citation.

Practical Implications and Future Directions

The study indicates that integrating attribution capabilities in LLMs can boost their validity and trustworthiness, especially in exploratory and verification-heavy domains. Moving forward, enhancing LLMs' proficiency in citing evidence for complex claim responses and advancing automatic evaluation models for attributability could elevate LLM usability in professional environments requiring high accuracy and reliability.

Conclusion

"Attribute or Abstain" underscores the potential and complexities of leveraging LLMs for long-document scenarios with a focus on evidence attribution. The insights into models' evidence-retrieval capabilities and positional biases offer a foundation for furthering research in developing more sophisticated AI systems capable of rigorous attribution, enhancing both trust and utility in academic and professional fields.