Attribution, Citation, and Quotation: A Survey of Evidence-based Text Generation with Large Language Models

1. Problem Statement

The rapid adoption of Large Language Models (LLMs) has raised significant concerns regarding their reliability, specifically their tendency to generate "hallucinations" (factual errors) and their reliance on static training data. To address this, researchers are developing evidence-based text generation, where LLM outputs are linked to supporting sources to ensure traceability and verifiability.

However, the field suffers from severe fragmentation:

• Inconsistent Terminology: Researchers use varied terms (e.g., citation, attribution, quotation, grounding) to describe similar or distinct concepts, leading to confusion.
• Isolated Evaluation: There is no unified benchmark or standardized evaluation framework, making it difficult to compare different approaches.
• Lack of Unified Taxonomy: Existing surveys focus on specific sub-topics (like RAG or hallucinations) rather than the holistic paradigm of evidence-based generation.

2. Methodology

The authors conducted a systematic mapping study following the PRISMA protocol to analyze the current research landscape.

• Data Collection: They queried nine major literature databases (including ACL Anthology, arXiv, IEEE Xplore) using a specific search string: ("large language model" OR "llm") AND ("citation" OR "attribution" OR "quote").
• Screening: From an initial pool of 1,040 records, they removed duplicates and applied inclusion criteria (English language, full text accessible, focus on LLM text generation with explicit evidence references).
• Final Dataset: 134 relevant papers published primarily between 2021 and early 2025 were selected for deep analysis.
• Annotation: The authors manually annotated these papers across multiple dimensions, extracting 300 distinct evaluation metrics, 19 frameworks, 231 datasets, and 11 benchmarks.
• Taxonomy Construction: They developed a multidimensional, facetted taxonomy to categorize the literature, allowing for multi-label assignments to capture the complexity of hybrid approaches.

3. Key Contributions

The paper makes three primary contributions to the field:

A. A Unified Taxonomy

The authors introduce the first comprehensive taxonomy for evidence-based text generation, categorizing papers along three independent dimensions:

1. Attribution Approach:
   • Parametric: Relies on knowledge encoded within model parameters (Pure LLM, Model-Centric, Data-Centric).
   • Non-Parametric: Relies on external evidence retrieved at inference time (Post-Retrieval, Post-Generation, In-Generation, In-Context).
2. Citation Characteristics:
   • Modality: Text (96% of studies), Graphs, Tables, Visuals.
   • Evidence Level: Granularity from tokens to full documents.
   • Style: In-line citations, citation reports, passages, narrative citations, highlight gradients, and quotes.
   • Visibility & Frequency: Whether citations are shown to users and if single or multiple citations are used per claim.
3. Task: Question Answering (QA), Grounded Text Generation, Summarization, Fact Verification, Citation Text Generation, and Related Work Generation.

B. Comprehensive Evaluation Analysis

The survey analyzes 300 evaluation metrics across seven dimensions:

• Core Dimensions: Attribution, Citation, and Correctness (essential for all systems).
• Contextual Dimensions: Linguistic Quality, Preservation, Relevance, and Retrieval (dependent on specific tasks).
• Findings: The field lacks standardization; only two frameworks (ALCE and G-Eval) and two benchmarks are frequently reused. Most studies rely on inference-based (NLI models) and lexical overlap metrics, while LLM-as-a-judge approaches are emerging but not yet dominant.

C. Identification of Trends and Gaps

The authors synthesize findings to highlight the evolution of the field and identify critical open challenges.

4. Key Results and Findings

• Research Growth: The field has exploded recently, with 75% of studies published after 2023. The number of papers surged from 28 in 2023 to 96 in 2024.
• Dominance of Non-Parametric Approaches: Post-Retrieval (RAG-style) approaches dominate the literature (58% of non-parametric methods), followed by In-Context (20%) and Post-Generation (18%). Parametric attribution remains underexplored (only 25 papers), with most focusing on evaluating pure LLMs rather than improving them.
• Modality Bias: The field is overwhelmingly text-centric (96% of studies). Evidence from graphs, tables, and visuals is largely unexplored, representing a major opportunity for multimodal research.
• Granularity Trends: Most studies cite evidence at the document or paragraph level. Fine-grained attribution (sentence or token level) is growing but remains rare.
• Evaluation Limitations:
  • Current automated metrics often fail to capture nuanced factual errors.
  • There is a trade-off between scalability (automated metrics) and semantic coverage (human evaluation).
  • "Correctness" is often evaluated via human judgment or LLM-as-a-judge, but standardized automated correctness metrics are still maturing.

5. Significance and Future Directions

This survey serves as a foundational resource for navigating the fragmented field of evidence-based LLM generation. Its significance lies in:

• Standardization: Providing a unified vocabulary and taxonomy to align future research.
• Benchmarking: Highlighting the urgent need for standardized evaluation frameworks to enable fair comparisons.
• Roadmap for Future Work: The authors outline four critical directions:
  1. Parametric and Hybrid Attribution: Moving beyond pure RAG to integrate internal model knowledge with external retrieval for better explainability and reduced retrieval dependency.
  2. Standardized Evaluation: Developing adaptable frameworks that cover all seven evaluation dimensions.
  3. Explainable Citations: Investigating why an LLM selects a specific source to improve trust and identify biases.
  4. Multimodal Evidence: Expanding beyond text to integrate and cite visual, tabular, and graph-based evidence.

In conclusion, the paper argues that while evidence-based text generation is a rapidly growing and promising field, it requires a shift from isolated, task-specific experiments toward a unified, standardized, and multimodal approach to truly enhance the reliability and trustworthiness of LLMs.