Citation Grounding (CG)

• Citation grounding is a methodology that verifies citations accurately support claims by linking specific text spans to their original sources.
• It employs techniques such as extractive span alignment, retrieval-augmented decoding, and graph-based provenance to ensure evidence fidelity.
• Empirical studies demonstrate that integrating fine-grained extraction and validation steps significantly improves citation precision and reduces hallucinations.

Citation grounding (CG) is the family of tasks concerned with ensuring that a generated citation faithfully supports the claim, sentence, span, or answer to which it is attached. Across current work, the concept is formalized at several granularities. FullCite defines an inline citation as $X=\{\text{doc\_id}:d,\ \text{snippet}:s\}$, where $d\in\mathcal{D}$ and $s$ is a contiguous, verbatim span from document $d$. CiteGuard treats CG as binary alignment over span–citation pairs $y_{ij}\in\{0,1\}$. Legal citation grounding verifies whether extracted citations exist and remain valid in a ground-truth citation graph. Agentic GraphRAG extends the notion beyond source support to trajectory-level provenance over cited and uncited traversal context (Yeginbergen et al., 5 Jun 2026, Choi et al., 15 Oct 2025, Ovcharov, 30 May 2026, Terrenzi et al., 14 May 2026).

1. Formal definitions and problem scope

In attributed question answering, FullCite formulates an answer as a sequence of claims $C_1,\dots,C_k$, each immediately followed by a citation $X_i$. Its citation grammar is

$S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$

with

$$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$

subject to two constraints: every claim must be followed by exactly one citation, and the snippet must appear verbatim in the chosen document. The resulting task is not merely to name a relevant document, but to bind each claim to a specific evidence span (Yeginbergen et al., 5 Jun 2026).

In scientific writing, CiteGuard defines CG over a generated document $D$ containing textual spans $d\in\mathcal{D}$0 and citations $d\in\mathcal{D}$1. For every observed pair $d\in\mathcal{D}$2, the task is to determine whether the citation is a faithful source for the claim in $d\in\mathcal{D}$3, represented by $d\in\mathcal{D}$4. The operational decision rule is retrieval-augmented: $d\in\mathcal{D}$5 where $d\in\mathcal{D}$6 is a retrieval score and $d\in\mathcal{D}$7 is an LLM-based alignment score (Choi et al., 15 Oct 2025).

Fine-grained grounding has also been defined as explicit extraction of supporting quotes. FRONT represents grounding as a sequence

$d\in\mathcal{D}$8

where each tuple contains a document index and an extractive segment drawn verbatim from that document. The answer sequence is then generated with citations restricted to document indices whose supporting segments appear in $d\in\mathcal{D}$9. Li et al. formulate a closely related scientific citation-generation problem as mapping each citation sentence $s$0 to one or more cited text spans $s$1 in the cited paper such that $s$2 (Huang et al., 2024, Li et al., 2023).

In legal reasoning, CG is additionally a metric. For a response $s$3 and legal citation graph $s$4,

$s$5

with $s$6 if no citations are extracted. This formulation is then decomposed into citation precision, citation relevance, and citation temporality, separating existence errors from contextual and temporal mismatch (Ovcharov, 30 May 2026).

Agentic GraphRAG broadens the scope further. Given a graph traversal $s$7, final citation faithfulness is defined not only by whether each cited node supports the answer, but by whether the cited set is necessary and sufficient relative to the full traversal and local neighborhoods. Terrenzi et al. therefore distinguish citation sufficiency from “provenance sufficiency,” where the induced subgraph over cited and visited nodes must be enough to recover model behavior (Terrenzi et al., 14 May 2026).

2. Grounding units and evidence granularity

The literature differs sharply in what counts as the atomic grounding unit. Some systems require contiguous verbatim spans from retrieved documents; others require extractive quotes, cited text spans from full papers, line intervals in code, statute nodes in a graph, or visited entities in a traversal trace. This variation is not superficial: it determines what can be verified mechanically, what can be scored automatically, and which failure modes remain latent (Li et al., 2023, Huang et al., 2024, Arafat, 13 Dec 2025, Ovcharov, 30 May 2026, Terrenzi et al., 14 May 2026).

Setting	Grounding unit	Verification condition
FullCite-style attributed QA	$s$8	snippet appears verbatim in the cited document
FRONT and CTS-based citation generation	extractive quote or cited text span	quote/span is drawn from the cited paper text
Code comprehension	$s$9	cited interval overlaps a retrieved chunk
Legal citation graphs	statute node with contextual and temporal checks	provision exists, is contextually appropriate, and was valid at the relevant date
Agentic GraphRAG	cited entities plus traversal context	faithfulness depends on cited and visited-but-uncited structure

Li et al. show that conditioning citation generation on cited text spans rather than on abstracts significantly $d$0 increases faithfulness, increases token overlap, and reduces hallucination. FRONT adopts an analogous fine-grained stance by requiring extractive supporting quotes before answer generation. In code, line-level citations permit strict interval-overlap checking. In legal settings, graph-based verification introduces a different granularity: the object being grounded is not a span of prose but a normalized law–article identifier, potentially indexed by date (Li et al., 2023, Huang et al., 2024, Arafat, 13 Dec 2025, Ovcharov, 30 May 2026).

A plausible implication is that “citation” is not a single data type across CG research. The field includes document identifiers, verbatim substrings, full-text evidence sentences, code line ranges, graph nodes, and traversal-induced provenance summaries. As a result, systems that appear similar at the output layer may be solving materially different grounding problems.

3. Methodological families

One line of work imposes structure directly during generation. FullCite implements three strategies: prompt-based generation, constrained decoding over a citation grammar compiled into a finite-state automaton, and posthoc span alignment. In constrained decoding, the automaton tracks whether decoding is in Claim, DOCID, or SNIPPET mode, restricts allowable tokens accordingly, and retries up to 3 times with temperature increased by $d$1 on failure. In posthoc alignment, a produced snippet $d$2 is replaced by the best matching verbatim span $d$3 if Jaccard token-set similarity satisfies $d$4, with $d$5 empirically (Yeginbergen et al., 5 Jun 2026).

A second line uses supervised or preference-based adaptation so that the model learns to self-cite. AGREE constructs grounded training examples from unlabeled queries using a pretrained NLI model $d$6, fine-tunes the base LLM to emit inline citations and identify unsupported statements, and then applies a test-time adaptation loop that re-retrieves on unsupported spans. XKD-Dial implements a four-stage progressive pipeline: multilingual adaptation, English citation-grounded SFT, bilingual citation-grounded SFT, and GRPO alignment with citation-aware rewards including factual consistency, entity overlap, citation attribution, fluency, length penalty, hallucination penalty, and explicit citation bonus or penalty (Ye et al., 2023, Pandya, 19 Mar 2026).

A third line validates citations after generation rather than entrusting the generative model alone. CiteGuard couples retrieval with fine-tuned validation: a query built from the textual span and citation metadata retrieves candidate documents, and a cross-encoder LLM scores whether a retrieved document supports the cited claim. Because the framework scores retrieved alternatives, it can identify alternative but valid citations in addition to rejecting spurious ones (Choi et al., 15 Oct 2025).

Production-oriented systems often combine citation prompting with retrieval and systems engineering constraints. The e-commerce “Cite Before You Speak” system uses in-context learning to generate answer sentences with bracketed citations in one pass and introduces Multi-UX-Inference, which stores citation-specific instructions on a separate attention-cache page so citation and non-citation UX variants can share a single prefill. In code comprehension, hybrid retrieval combines BM25, BGE dense embeddings, and Neo4j import-graph expansion, after which mechanical citation verification checks that every cited line range overlaps a retrieved chunk; if no valid citations are emitted, an auto-cite fallback inserts the highest-scoring retrieved chunk (Zeng et al., 5 Mar 2025, Arafat, 13 Dec 2025).

4. Evaluation frameworks

FullCite evaluates grounding along three dimensions. Document-level correctness uses the gold and predicted supporting document sets $d$7 and $d$8, with precision $d$9, recall $y_{ij}\in\{0,1\}$0, and

$y_{ij}\in\{0,1\}$1

Evidence span identification uses token-level F1 between predicted and gold snippets in the same document, averaged over claims. Claim-citation faithfulness is measured by cosine similarity between embeddings of the claim and citation using Sentence-Transformer all-MiniLM-L6-v2, by GPT-5.4 support and relevance judgments on a 1–5 scale, and by human evaluation on the same rubric (Yeginbergen et al., 5 Jun 2026).

Other settings broaden the metric set. The e-commerce system defines Claim Grounding Rate, Correct Citation Rate, Perfect Sentence Rate, Sentence with Citation Rate, and Evidence Utilization Rate, all using an NLI judge $y_{ij}\in\{0,1\}$2. XKD-Dial reports BLEU, ROUGE, BERTScore, FactScore, Citation-F1, and hallucination rate, and supplements outcome metrics with cross-attention alignment, Integrated Gradients attribution, and occlusion-based causal grounding. These explainability analyses aim to reveal how citation behavior is learned, not only whether it is learned (Zeng et al., 5 Mar 2025, Pandya, 19 Mar 2026).

Legal citation grounding decomposes quality into citation precision, citation relevance, and citation temporality. The derived Citation Quality score penalizes non-existent citations, contextually inappropriate citations, and citations that were not valid at the relevant date. LegalCiteBench, by contrast, studies closed-book authority generation with citation retrieval, citation completion, citation error detection, case matching, and case verification/correction, and introduces Misleading Answer Rate as a measure of concrete but incorrect outputs under low task score (Ovcharov, 30 May 2026, Chen et al., 11 May 2026).

Trajectory-sensitive systems require different diagnostics. Agentic GraphRAG measures necessity through $y_{ij}\in\{0,1\}$3, contrasts it with random removal $y_{ij}\in\{0,1\}$4, tests sufficiency through $y_{ij}\in\{0,1\}$5, and tracks

$y_{ij}\in\{0,1\}$6

These metrics explicitly evaluate whether final citations alone can reconstruct behavior or whether uncited traversal context remains causally active (Terrenzi et al., 14 May 2026).

5. Empirical findings across domains

In attributed QA and scientific writing, a recurrent pattern is strong document identification but weak evidence localization. FullCite reports that LLMs are generally effective at identifying relevant documents, with Doc-F1 often in the 50–90% range, but struggle to identify precise supporting spans, with Snippet-F1 often below 30%. On ASQA, prompt-based generation with Qwen3-8B yields $y_{ij}\in\{0,1\}$7 for document, snippet, and similarity, whereas FullCite posthoc yields $y_{ij}\in\{0,1\}$8; the ASQA snippet-F1 jump from $y_{ij}\in\{0,1\}$9 to $C_1,\dots,C_k$0 is a central result. FullCite also reports that citation enforcement improves downstream QA for BioASQ factoid questions from $C_1,\dots,C_k$1 to $C_1,\dots,C_k$2 ROUGE-L. FRONT, on the ALCE benchmark with LLaMA-2-7B, reports citation F1 scores of $C_1,\dots,C_k$3 on ASQA, $C_1,\dots,C_k$4 on ELI5, and $C_1,\dots,C_k$5 on QAMPARI, corresponding to an average $C_1,\dots,C_k$6 improvement in citation quality across all datasets and surpassing ChatGPT on those citation-F1 measurements. CiteGuard raises CiteME accuracy from $C_1,\dots,C_k$7 for LLM-as-a-Judge to $C_1,\dots,C_k$8, compared with $C_1,\dots,C_k$9 human inter-annotator accuracy (Yeginbergen et al., 5 Jun 2026, Huang et al., 2024, Choi et al., 15 Oct 2025).

Training-based self-grounding also improves classical grounding metrics. AGREE reports citation-recall gains of 15–40 percentage points and citation-precision gains of 20–50 percentage points over prompting-based and post-hoc approaches across five datasets, with test-time adaptation adding 2–6 points in recall and 1–7 points in precision; on NQ, TTA also yields $X_i$0 EM. XKD-Dial reports that citation-grounded SFT reduces hallucination to $X_i$1 for encoder–decoder models from Stage 2 onward, that both Flan-T5-Base and Flan-T5-Large reach English Citation-F1 $X_i$2 after Stage 2, and that Stage 3 to Stage 4 deltas are negligible, below $X_i$3 in Citation-F1 and FactScore (Ye et al., 2023, Pandya, 19 Mar 2026).

In production QA, citation prompting can improve both grounding metrics and user-facing outcomes. In the e-commerce CSA setting, CGR rises on synthetic data from $X_i$4 for Vanilla to $X_i$5 for Guided and $X_i$6 for Citation. On real-world data, CGR rises from $X_i$7 for Guided to $X_i$8 for Citation, a relative improvement of approximately $X_i$9. The same system reports $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$0, $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$1, $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$2, and refusal signals for no-relevant-evidence queries $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$3 of the time under citation prompting versus $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$4 under Guided. Large-scale online A/B tests over 10 M customers for two weeks report customer-engagement gains of $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$5, depending on the UX (Zeng et al., 5 Mar 2025).

Scientific citation-text generation shows a similar gain from finer evidence. Li et al. report that abstract-only conditioning gives ROUGE-L recall $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$6 and ANLI $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$7, whereas CTS-based inputs push ROUGE-L recall above $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$8 and ANLI above $S \to \text{Claim}\;\text{CITE}\;S \mid \epsilon,$9. Their keyword-based CTS retrieval achieves input ROUGE-L recall $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$0 and downstream BLEU $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$1, compared with abstract BLEU $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$2. Human evaluation for RAG-FiD gives Overall scores of $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$3 for Abstract, $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$4 for Context, $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$5 for Oracle, and $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$6 for Keyword (Li et al., 2023).

In code comprehension, the strongest reported numbers come from hybrid retrieval plus strict verification. Across 30 Python repositories and 180 developer queries, the system reports $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$7 citation accuracy with zero hallucinations. Sparse-only and dense-only modes yield citation accuracy of $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$8 and $$\text{CITE} \to \{\text{doc\_id:}\;ID,\ \text{snippet:}\;SNIP\},$$9, respectively, while the hybrid system reaches $D$0, a $D$1 point gain over sparse and $D$2 over dense. For 112 cross-file questions, graph expansion raises cross-file citation completeness from $D$3 to $D$4, discovering on average 11.8 new neighbor chunks per query (Arafat, 13 Dec 2025).

Legal results are bifurcated between grounded verification and closed-book generation. The Ukrainian legal citation-graph framework reports mean CG in the $D$5 range across five systems, corresponding to $D$6 hallucinated citations. On a CG-DPO dataset of 2,244 court decisions, a Qwen2.5-7B-Instruct model fine-tuned with LoRA reaches mean validation accuracy $D$7, reward margin $D$8, and standard deviation below $D$9 percentage points across 3 seeds in distinguishing correct from corrupted citations. LegalCiteBench, which tests closed-book citation recovery on approximately 24K instances and 21 LLMs, reports that even the strongest models score below $d\in\mathcal{D}$00 on citation retrieval and completion, and that Misleading Answer Rates exceed $d\in\mathcal{D}$01 for 20 of 21 evaluated models on retrieval-heavy tasks (Ovcharov, 30 May 2026, Chen et al., 11 May 2026).

6. Limitations, misconceptions, and open problems

A persistent misconception is that citation syntax is equivalent to citation grounding. XKD-Dial explicitly shows otherwise: decoder-only models often learn citation format without true causal grounding, and Citation-F1 must therefore be complemented by occlusion tests. The same paper reports that encoder–decoder models achieve high causal-grounding scores, approximately $d\in\mathcal{D}$02, whereas decoder-only models can drop to $d\in\mathcal{D}$03, indicating that citation markers may be emitted without consulting the source text (Pandya, 19 Mar 2026).

A second misconception is that support from final citations is sufficient to explain system behavior. Agentic GraphRAG directly contests that assumption. On the 25-question testbed after excluding LLM-solvable items, cited removal in Agentic GraphRAG drops accuracy from $d\in\mathcal{D}$04 to $d\in\mathcal{D}$05, showing necessity, but full isolation to cited entities alone still degrades performance strongly. In the Evidence-First variant, full isolation drops accuracy from $d\in\mathcal{D}$06 to $d\in\mathcal{D}$07, while text-only isolation for Agentic GraphRAG drops it only from $d\in\mathcal{D}$08 to $d\in\mathcal{D}$09. Removing visited-but-uncited entities lowers Agentic accuracy from $d\in\mathcal{D}$10 to $d\in\mathcal{D}$11. These results indicate that faithful provenance may require cited evidence, uncited traversal context, and surrounding graph structure together (Terrenzi et al., 14 May 2026).

Span identification remains a principal bottleneck even when document retrieval is strong. FullCite identifies precise evidence extraction as the main gap, notes BioASQ snippet-F1 around $d\in\mathcal{D}$12 even in the best setting, and reports a primacy bias in which $d\in\mathcal{D}$13 of BioASQ citations target only the first two of five documents. It also notes that yes/no questions often omit citations entirely, inflating ungrounded QA metrics. CiteGuard likewise questions reliance on LLM-as-a-Judge alone and motivates retrieval-aware validation as a way to reduce hallucinated or spurious citations (Yeginbergen et al., 5 Jun 2026, Choi et al., 15 Oct 2025).

Closed-book legal citation generation exposes a distinct failure mode: calibration can improve without improving knowledge. LegalCiteBench reports that explicit uncertainty instructions reduce some confident fabrication and lower Misleading Answer Rate in prompt-only abstention experiments, but do not improve citation correctness. The graph-based legal CG work further notes that $d\in\mathcal{D}$14 of “hallucinations” flagged by its metric were real statutes that courts rarely cite, and that current temporality checks do not yet use versioned force-of-law data, motivating integration of a legislative API and broader graph coverage (Chen et al., 11 May 2026, Ovcharov, 30 May 2026).

The future directions proposed across the literature are correspondingly concrete. FullCite recommends supervised span-locator modules, hybrid retrieval-generation that retrieves candidate spans first, explicit instruction tuning for joint document-plus-span grounding, and finer-grained evaluation such as position-based citation recall. AGREE proposes human-annotated partial-support examples, RLHF with grounded-response preferences, and extension to multilingual corpora and long-form generation. The e-commerce CSA work points to stronger retrievers, dynamic source selection, and multimodal UXs. Code-grounded systems propose richer structural analyses, including call graphs and data flow, as well as citation anchors less brittle than raw line numbers. Taken together, these proposals suggest that CG is shifting from a formatting problem toward a broader evidence-engineering problem in which retrieval, alignment, verification, and provenance all remain active research fronts (Yeginbergen et al., 5 Jun 2026, Ye et al., 2023, Zeng et al., 5 Mar 2025, Arafat, 13 Dec 2025).