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

Conference: ACL 2026 arXiv: 2508.15396 Code: https://github.com/faerber-lab/AttributeCiteQuote Area: Survey / NLP Keywords: Evidence-based text generation, citation attribution, LLM trustworthiness, evaluation framework, RAG

TL;DR

This paper presents a systematic survey of 134 papers on evidence-based text generation with LLMs, proposing for the first time a unified taxonomy (attribution approach × citation characteristics × task), analyzing 300 evaluation metrics organized into seven dimensions and six method categories, and providing a panoramic reference framework for this fragmented field.

Background & Motivation

Background: LLMs face trustworthiness challenges such as hallucination and knowledge limitations, driving increasing research interest in "evidence-based text generation"—enabling LLM outputs to be traceable to supporting evidence. However, the field is highly fragmented: some work uses the term "citation" (75% of papers), others use "attribution" (62%), and still others use "quotation" (13%), with each subcommunity maintaining isolated evaluation practices.

Limitations of Prior Work: (1) The absence of unified terminology and a classification system makes it difficult for researchers to situate their own work; (2) evaluation standards are inconsistent—300 metrics exist, yet only 2 frameworks (ALCE and G-Eval) are reused across multiple papers; (3) RAG, though popular, represents only one of seven relevant methods, and excessive focus on RAG risks overlooking other important approaches.

Key Challenge: Rapidly growing research interest (the number of papers in 2024 is 3.4× that of 2023, with 75% published after 2023) versus the lack of a unified perspective for integrating and comparing different methods.

Goal: To provide the first systematic survey dedicated to evidence-based text generation with LLMs, establish a unified taxonomy, analyze evaluation practices, and identify research trends and future directions.

Key Insight: A PRISMA-protocol systematic mapping study, screening 134 relevant papers from 805 deduplicated papers, with a faceted classification approach used to construct the taxonomy.

Core Idea: Unifying "citation," "attribution," and "quotation" under the paradigm of "evidence-based text generation," and providing a systematic perspective on this fragmented field through a three-dimensional taxonomy and a seven-dimensional evaluation framework.

Survey Framework / Taxonomy

Overall Architecture

The paper proposes a three-dimensional, orthogonal taxonomy capturing core design choices in evidence-based text generation:

• Dimension 1: Attribution Approach — how content is linked to evidence (parametric vs. non-parametric)
• Dimension 2: Citation Characteristics — the form and presentation of evidence (modality, granularity, style, visibility, frequency)
• Dimension 3: Task — application scenarios (six task categories including QA, grounded generation, summarization, etc.)

Key Designs

1. Attribution Approach:

   • Function: Describes how LLMs associate generated content with supporting evidence.
   • Core taxonomy: Parametric (25 papers)—pure LLM (leveraging existing capabilities, 72% of parametric), model-centric (modifying architecture/training), data-centric (curating/augmenting data). Non-parametric (126 papers)—post-retrieval (58%, RAG being representative), post-generation (18%, generating first then retrieving evidence), in-generation (4%, dynamically deciding whether retrieval is needed), in-context (20%, user directly provides evidence).
   • Key Findings: Parametric attribution is severely underexplored, with model-centric and data-centric approaches receiving particularly little attention. Among non-parametric methods, post-retrieval dominates, but in-generation attribution (e.g., Self-RAG) represents a promising yet undervalued emerging direction.

2. Citation Characteristics:

   • Function: Describes the specific presentation of evidence.
   • Core taxonomy: Citation modality—text 96%, figures, tables, visuals. Evidence granularity—document-level 43%, paragraph-level 40%, sentence-level 12%, token-level 2%. Citation style—inline citation 62%, citation report, paragraph display, narrative citation, highlight gradient, quotation. Visibility—final response 91% vs. intermediate text. Frequency—multiple citations 64% vs. single citation.
   • Key Findings: Non-textual evidence modalities (figures, tables, images) remain severely underexplored (only 4%); fine-grained evidence (sentence-level, token-level), though a small share, is growing faster.

3. Task Distribution:

   • Function: Maps the application landscape of the field.
   • Core taxonomy: QA and grounded text generation are dominant tasks; summarization and fact verification occupy a middle tier; citation text generation and related work generation are emerging tasks.
   • Key Findings: Evaluation practices have been developed primarily around QA tasks and may not be appropriate for emerging tasks (e.g., citation text generation requires more evaluation of citation selection reasoning).

LLM Integration

Training (used in 45% of papers): Dominated by supervised fine-tuning, primarily to improve attribution behavior. Pre-training is less common. Prompting (used in 78% of papers): Dominated by zero/few-shot prompting. Strategies specifically targeting citation behavior include chain-of-citation, chain-of-quote, and conflict-aware prompting.

Literature Analysis / Coverage

Evaluation Metric Framework

300 metrics organized into seven evaluation dimensions

Evaluation Dimension	When to Use	Primary Methods	Representative Metrics (reuse count)
Attribution	When annotated evidence is unavailable	Primarily NLI	Citation NLI P/R/F1 (33/33/16), Auto-AIS (11), FActScore (7)
Citation	When annotated evidence is available	Primarily retrieval-based	Citation Retrieval P/R/F1 (6/6/5), Citation Accuracy (2)
Correctness	Always required	Lexical overlap / NLI	Exact Match (12), BLEU-N (5), Claim Recall (17)
Language Quality	When model is modified	LLM-as-Judge	G-Eval Fluency (4), MAUVE (21), Perplexity (4)
Preservation	For post-generation attribution	Lexical overlap	Preservation-Levenshtein (3), F1-AP (2)
Relevance	User-facing scenarios	LLM-as-Judge	G-Eval Relevance (3), RAGAS (2)
Retrieval	For non-parametric attribution	Retrieval metrics	P@k (4), R@k (4), MRR (3)

Evaluation Guidelines

Dimension Category	When to Evaluate	Notes
Core Dimensions	Attribution or Citation + Correctness	Correctness should always be evaluated; attribution and citation are chosen based on evidence availability
Contextual Dimensions	Language Quality, Preservation, Relevance, Retrieval	Depends on task design and system architecture

Key Findings

• Only 2 frameworks (ALCE and G-Eval) and 2 benchmarks are reused across multiple papers, indicating a severe lack of evaluation standardization.
• 19 frameworks, 11 benchmarks, and 231 datasets are identified across the 134 papers.
• Text accounts for 96% of citation modalities; multimodal evidence remains virtually unexplored.
• Parametric attribution, despite being critical for understanding models' internal knowledge and data provenance, is severely neglected.
• Human evaluation still dominates the correctness dimension, reflecting the limitations of automatic metrics in capturing semantic errors.

Highlights & Insights

• Unifying "citation," "attribution," and "quotation" under "evidence-based text generation" constitutes an important conceptual contribution, resolving longstanding terminological confusion.
• The seven-dimensional evaluation guideline (Table 1) provides practitioners with clear metric selection recommendations—the distinction between core and contextual dimensions is highly practical.
• The tripartite classification of parametric attribution (pure LLM / model-centric / data-centric) is more nuanced than previous binary distinctions.
• In-generation attribution is identified as a promising yet undervalued direction—accounting for only 4% of papers but representing a trend toward tighter integration of retrieval and generation.
• The paper notes that LLM citation behavior may exhibit biases analogous to those of human authors, calling for research into the interpretability of LLM citation reasoning.

Limitations & Future Work

• A single search string may miss some relevant studies (sensitivity analysis shows only 4% additional papers found).
• Coverage is limited to English-language papers, potentially underrepresenting non-English research.
• Manual screening and classification inevitably introduce some degree of subjectivity.
• Four major future directions: (1) deeper investigation of parametric and hybrid attribution; (2) standardized evaluation frameworks (currently only 2 frameworks are reused among 300 metrics); (3) interpretable citation reasoning—understanding why LLMs select particular sources; (4) multimodal evidence support—extending beyond the current 96% text dominance to figures, tables, and images.

Related Work & Insights

• vs. Li et al. (2023a): The only prior related survey, but now severely outdated (75%+ of papers were published after it) and does not cover the full paradigm.
• vs. Huang & Chang (2024): A position paper that only emphasizes the importance of citation without providing a systematic survey.
• vs. RAG surveys: RAG surveys cover only post-retrieval, one of the seven attribution approaches covered in this paper.
• vs. hallucination/grounding surveys: Those surveys focus on different aspects; this paper is specifically concerned with evidence generation rather than detection.

Rating

• Novelty: ⭐⭐⭐⭐ First comprehensive unified taxonomy for evidence-based text generation; the three-dimensional taxonomy is well-designed.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Coverage of 134 papers, 300 metrics, 19 frameworks, 231 datasets, and 11 benchmarks is highly comprehensive.
• Writing Quality: ⭐⭐⭐⭐⭐ Clear structure, balanced multi-dimensional analysis, with concise "Takeaways" summarizing each section.
• Value: ⭐⭐⭐⭐⭐ A panoramic synthesis of a rapidly growing yet fragmented field, providing important reference value for both researchers and practitioners.

Related Papers

• [ACL 2026] Capabilities and Evaluation Biases of Large Language Models in Classical Chinese Poetry Generation: A Case Study on Tang Poetry
• [ACL 2026] Closing the Modality Reasoning Gap for Speech Large Language Models
• [AAAI 2026] Structured Language Generation Model: Loss Calibration and Formatted Decoding for Efficient Text
• [ACL 2026] AnchorMem: Anchored Facts with Associative Contexts for Building Memory in Large Language Models
• [ACL 2026] E2EDev: Benchmarking Large Language Models in End-to-End Software Development Task