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 systematically reviews 134 papers on evidence-based text generation for LLMs. It proposes the first unified taxonomy (Attribution Mechanism × Citation Features × Task), analyzes 300 evaluation metrics categorized into seven dimensions and six methods, and provides a panoramic reference framework for this fragmented field.

Background & Motivation¶

Background: LLMs face trustworthiness challenges such as hallucinations and knowledge limitations. Increasing research focuses on "evidence-based text generation"—ensuring LLM outputs are traceable to supporting evidence. However, the field is highly fragmented: various terms are used, such as "citation" (75% of papers), "attribution" (62%), or "quotation" (13%), with isolated evaluation practices.

Limitations of Prior Work: (1) Lack of unified terminology and classification systems, making it difficult for researchers to position their work; (2) Inconsistent evaluation standards—among 300 metrics, only two frameworks (ALCE, G-Eval) are reused across multiple papers; (3) Although RAG is popular, it is only one of seven relevant methods, and over-focusing on RAG misses other critical approaches.

Key Challenge: Rapidly growing research interest (the number of papers in 2024 is 3.4x that of 2023, with 75% published after 2023) vs. the lack of a unified perspective to integrate and compare different methods.

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

Key Insight: A systematic mapping study was conducted following the PRISMA protocol, screening 134 relevant papers from 805 deduplicated candidates, using a faceted classification method to construct the taxonomy.

Core Idea: Unify "citation," "attribution," and "quotation" into an "evidence-based text generation" paradigm, providing a systematic perspective through a three-dimensional taxonomy and a seven-dimensional evaluation framework.

Method¶

Overall Architecture¶

A systematic mapping study was performed per the PRISMA protocol: 134 relevant works were screened from 805 deduplicated papers, followed by manual encoding using a faceted classification method. The core output is a 3D independent taxonomy—combining "Attribution Mechanism (how content relates to evidence) × Citation Features (the form of the evidence) × Task (application scenario)." Any work in evidence-based text generation can be positioned within this cube. A cross-cutting perspective of LLM integration (Training vs. Prompting) is overlaid to answer how models acquire attribution capabilities.

Key Designs¶

1. Attribution Mechanism: Dichotomy and Subdivision of Parametric vs. Non-parametric. This dimension characterizes the fundamental path through which LLMs associate generated content with supporting evidence. Parametric (25 papers) involves evidence entering model weights, subdivided into Pure LLM (leveraging existing capabilities, 72% of parametric), Model-centric (architecture/training changes), and Data-centric (data curation/augmentation). Non-parametric (126 papers) keeps evidence outside the weights, categorized by retrieval timing: Post-retrieval (58%, e.g., RAG), Post-generation (18%, generating then finding evidence), In-generation (4%, dynamic decision to retrieve), and In-context (20%, evidence directly in the prompt). This three-way split is finer than the simple "RAG / Non-RAG" distinction and reveals field imbalances: parametric methods are largely neglected, while in-generation attribution (e.g., Self-RAG), though only 4%, represents the frontier of tight coupling.

2. Citation Features: Five facets characterizing evidence appearance. The same evidence can be presented differently. These facets include Citation Modality (96% text, with charts/visuals nearly blank), Evidence Level (Document 43%, Paragraph 40%, Sentence 12%, Token 2%), Citation Style (In-line 62%, plus reports, snippets, narrative, highlight gradients, quotations, etc.), Visibility (Final response 91% vs. intermediate text), and Frequency (Multiple 64% vs. Single). These reveal two structural gaps: non-text modalities are severely underdeveloped (4%), while fine-grained evidence (sentence/token level) is the fastest-growing segment, indicating a trend toward precise traceability.

3. Task: Mapping six application scenarios. The third dimension maps tasks, finding QA and Grounded Text Generation as dominant, with Summarization and Fact-checking in the middle, and Citation Text Generation or Related Work Generation as emerging tasks. This reveals path dependency in evaluation: existing metrics are largely designed for QA and might not apply to emerging tasks that require reasoning about "why" a citation was chosen.

Additionally, LLM Integration methods are categorized into Training (45% of papers, primarily SFT to improve attribution behavior) and Prompting (78%, primarily zero/few-shot, with specialized strategies like chain-of-citation or conflict-aware prompting).

Key Experimental Results¶

Literature Analysis/Coverage¶

Evaluation Metric System¶

300 metrics classified into seven evaluation dimensions

Evaluation Dimension	When to Use	Primary Method	Representative Metrics (Reuse Count)
Attribution	No ground-truth evidence	NLI-based	Citation NLI P/R/F1 (33/33/16), Auto-AIS (11), FActScore (7)
Citation	Ground-truth evidence exists	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	Post-generation attribution	Lexical overlap	Preservation-Levenshtein (3), F1-AP (2)
Relevance	User scenarios	LLM-as-Judge	G-Eval Relevance (3), RAGAS (2)
Retrieval	Non-parametric attribution	Retrieval metrics	P@k (4), R@k (4), MRR (3)

Evaluation Guidelines¶

Dimension Category	When to Evaluate	Description
Core Dimensions	Attribution or Citation + Correctness	Correctness should always be evaluated; Attribution vs. Citation depends on evidence availability.
Contextual Dimensions	Language Quality, Preservation, Relevance, Retrieval	Depends on task design and system architecture.

Key Findings¶

Only two frameworks (ALCE, G-Eval) and two benchmarks are reused across multiple papers, indicating severe lack of standardization.
Identified 19 frameworks, 11 benchmarks, and 231 datasets among 134 papers.
Text accounts for 96% of citation modalities; multimodal evidence is almost nonexistent.
Parametric attribution is crucial for understanding internal knowledge and data provenance but is significantly neglected.
Human evaluation still dominates the correctness dimension, reflecting the limitations of automated metrics in capturing semantic errors.

Highlights & Insights¶

Unifying "citation," "attribution," and "quotation" as "evidence-based text generation" is a significant conceptual contribution, resolving long-standing terminological confusion.
The seven-dimension evaluation guide (Table 1) provides clear metric selection advice for practitioners—the distinction between core and contextual dimensions is highly practical.
The three-way split for parametric attribution (Pure LLM / Model-centric / Data-centric) is more refined than previous binary classifications.
In-generation attribution is identified as a promising but undervalued direction (4%), representing the trend of tighter integration between retrieval and generation.
Notes that citation behavior may exhibit biases similar to 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 showed only 4% additional findings).
Only covers English papers, potentially underestimating non-English research.
Manual screening and classification inevitably introduce some subjectivity.
Four Future Directions: (1) Deep research into parametric and hybrid attribution; (2) Standardized evaluation frameworks (only 2 of 300 metrics/frameworks are reused); (3) Explainable citation reasoning—understanding why LLMs select specific sources; (4) Multimodal evidence support—extending from 96% text to charts, tables, and images.

vs. Li et al. (2023a): The only prior relevant survey, but largely outdated (75%+ papers published after it), and does not cover the full paradigm.
vs. Huang & Chang (2024): A position paper that emphasizes citation importance but lacks a systematic survey.
vs. RAG surveys: RAG surveys only cover post-retrieval; this paper covers seven attribution mechanisms.
vs. Hallucination/Grounding surveys: Focuses on different aspects; this paper concentrates on evidence generation rather than detection.

Rating¶

Novelty: ⭐⭐⭐⭐ First comprehensive unified classification for evidence-based text generation; 3D taxonomy is well-designed.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Extremely comprehensive coverage of 134 papers, 300 metrics, 19 frameworks, 231 datasets, and 11 benchmarks.
Writing Quality: ⭐⭐⭐⭐⭐ Clear structure, balanced multi-dimensional analysis, and concise "Takeaways" for each section.
Value: ⭐⭐⭐⭐⭐ A panoramic synthesis of a rapidly growing but fragmented field, providing significant reference value for researchers and practitioners.