Generation-Time Citation (G-Cite)

• Generation-Time Citation (G-Cite) is a paradigm that integrates real-time citation insertion with text generation for immediate, verifiable source attribution.
• It contrasts with post-hoc citation methods by merging content production and citation retrieval in a single left-to-right process, ensuring stronger evidence links.
• Research in G-Cite spans LLM attribution accuracy and multi-generation scientometrics, leveraging techniques like contextual embeddings, citation alignment, and knowledge graphs.

Generation-Time Citation (G-Cite) denotes, in recent large-language-model attribution literature, the paradigm in which a model produces an answer and its citations in a single left-to-right generation process, so that citations are integrated into the text as it is written rather than attached afterward (Saxena et al., 25 Sep 2025). Its central purpose is attribution: generated statements are paired with human-verifiable sources so that users can trace claims back to evidence. In the literature represented here, the same label also appears in an older scientometric sense concerned with citation generations, citation cascades, and multi-generation knowledge ancestry rather than online attribution. The two usages are related by a common concern with how knowledge is linked across sources, but they address different technical objects and evaluation questions (Min et al., 2020, Min et al., 2021).

1. Definition, scope, and competing usages

In the attribution setting, G-Cite is defined by joint generation of content and citations. The model makes citation decisions locally from the content generated so far and any retrieved evidence, such as passages, documents, or sentences. This contrasts with Post-hoc Citation (P-Cite), in which an answer is drafted first and citations are attached or verified in a subsequent step (Saxena et al., 25 Sep 2025). The distinction is operational: G-Cite merges content production and attribution into one pass, whereas P-Cite separates them.

A separate usage appears in citation-network analysis. There, “generation-time citation” is tied to direct and indirect citation generations, citation cascades, and references of references. A citation cascade is a rooted directed acyclic graph whose root is the focal publication, whose children are direct and indirect citing publications, and whose generations are defined by shortest-path distance from the root. A backward analogue traces a paper’s references, then references of those references, and so on, as a “reference cascade” or “knowledge ancestry” (Min et al., 2020, Min et al., 2021).

These usages should not be conflated. In recent LLM work, G-Cite is a decoding-time attribution paradigm. In scientometrics, the same label is linked to multi-generational citation influence. The latter literature nonetheless provides useful conceptual background: it shows that indirect citation relations can remain meaningful for several generations. One study reports that two to four generations are generally adequate to trace a work’s scientific impact, while another finds that backward citation generations within six generations maintain Jaccard similarity above 0.2 in two APS sub-fields (Min et al., 2020, Min et al., 2021).

2. Core attribution problem in LLM-based G-Cite

The immediate motivation for G-Cite is that attribution enhances the credibility of LLM outputs by adding citations to generated sentences, enabling users to trace source material and verify reliability. The core difficulty is that existing instruction-tuned attributed LLMs often do not interpret citation markers such as $[i]$ as semantically grounded references. Instead, the markers behave as generic placeholders, weakening the link between generated claims and retrieved evidence (Yu et al., 19 Nov 2025).

This placeholder failure has several consequences. The literature describes weak semantic links between in-text citation markers and underlying content, disjointed or incorrect citation placement, and degraded fluency and correctness caused by naive insertion of disconnected citation tokens. In this diagnosis, the problem is not simply whether a citation appears, but whether the model understands what the symbol is supposed to point to during generation (Yu et al., 19 Nov 2025).

Related work on citation text generation reaches an analogous conclusion from a different angle. Conditioning only on the cited paper’s abstract is often insufficient, because gold citation texts frequently refer to material in the body rather than the abstract. Models trained on abstract-only inputs learn to hallucinate, and both models and human judges assign lower relevance and factuality scores when important supporting context is missing. The proposed alternative is to condition on cited text spans (CTS), namely the specific spans in the cited paper on which the citation text is based (Li et al., 2023).

This suggests a general principle across G-Cite and citation text generation: attribution quality depends on how directly the model’s internal citation representation is tied to the actual evidential content that the citation is supposed to denote.

3. Mechanisms for grounding citation generation

C$^2$-Cite is an explicit attempt to convert citation symbols from passive placeholders into active, semantically grounded knowledge pointers. Its first mechanism is contextual-aware embedding. Citation markers are normalized from multi-token form such as $[i]$ to a single special token $\langle c_i \rangle$, and the embedding of that token is replaced with a semantic embedding computed from the referenced document:

$$c_i = \mathcal{F}_{\text{contextual}}(e_1^{(i)}, e_2^{(i)}, \dots, e_N^{(i)}).$$

Here $e_j^{(i)}$ is the embedding of the $j$-th token in the referenced document $d_{[i]}$, and mean-pooling is used for $\mathcal{F}_{\text{contextual}}$. The intended effect is that citation tokens “carry” the meaning of their respective sources (Yu et al., 19 Nov 2025).

Its second mechanism is contextual citation alignment. During decoding, a citation router function, implemented as a binary classifier, predicts whether the current output position is a citation:

$$\hat{r}_t = \text{Softmax}(\mathbf{W}_r \cdot \mathbf{h}_t^{\top}).$$

For citation positions, the hidden state is mapped into the semantic space of contextual citation embeddings,

$^2$0

and then aligned with candidate citation embeddings through

$^2$1

The framework further adds contextual attentive augment, in which attention to previously generated citation tokens is strengthened with a locality bias:

$^2$2

Training uses a joint objective combining standard token loss, citation alignment loss, router loss, and attentive augmentation loss (Yu et al., 19 Nov 2025).

A complementary grounding line replaces abstract-only evidence with cited text spans. CTS are retrieved by distant labeling with a lexical-overlap criterion using the gold citation text at training time:

$^2$3

At test time, because the gold citation text is unavailable, the paper proposes a human-in-the-loop keyword-based retrieval setup, with T5 for keyword extraction and Dense Passage Retrieval with SciBERT-based encoders for sentence retrieval from the cited paper (Li et al., 2023).

Another grounding strategy is knowledge-graph-guided citation generation. KG-CTG constructs knowledge graphs from source and cited abstracts with PL-Marker, concatenates graph triples with textual components such as abstracts, introduction, and conclusion, and fine-tunes LLaMA-7B, Alpaca, and Vicuna for citation generation. The paper frames the KG as explicit structural context that helps the model learn relations between the citing and cited papers (Anand et al., 2024).

4. Benchmarks, task formulations, and evaluation protocols

Evaluation of G-Cite is fragmented because the field spans attribution, citation text generation, and citation prediction. One attribution-centered comparison evaluates G-Cite and P-Cite across four datasets—ALCE, LongCite (LongBench-Cite), REASONS, and FEVER—and reports citation correctness, citation precision, citation recall, coverage, and latency. Citation correctness is defined as the harmonic mean of precision and recall:

$^2$4

This protocol is intended to expose trade-offs among correctness, coverage, and speed rather than only raw citation presence (Saxena et al., 25 Sep 2025).

CiteBench standardizes scientific citation text generation by formalizing the task as generation of citation text $^2$5 from a citing document $^2$6, a set of cited documents $^2$7, and contextual windows $^2$8. It unifies four previously divergent datasets—ABURAED, XING, LU, and CHEN—and evaluates models with ROUGE, BERTScore, citation intent labeling, CORWA discourse tagging, and KL-divergence over intent and discourse distributions (Funkquist et al., 2022).

CiteRAG extends evaluation to academic citation prediction under a retrieval-augmented generation setting. It defines two tasks. Task 1 is coarse-grained list-specific citation prediction, where the system predicts a full reference list from a query paper’s title and abstract. Task 2 is fine-grained position-specific citation prediction, where the system predicts the correct citation for a reference placeholder using surrounding context. The benchmark contains 7,267 instances for Task 1 and 8,541 instances for Task 2, supported by a three-level corpus of 554,719 papers (Zheng et al., 21 Jan 2026).

Benchmark or setting	Focus	Key details
ALCE-based attribution evaluation	Generation-time and post-hoc attribution	ALCE, LongCite, REASONS, FEVER; correctness, precision, recall, coverage, latency
CiteBench	Citation text generation	Unifies ABURAED, XING, LU, CHEN; ROUGE, BERTScore, discourse metrics
CiteRAG	Citation prediction with RAG	7,267 Task 1 instances; 8,541 Task 2 instances; 554,719-paper corpus

The metrics used in CiteRAG are broader than standard ranking scores. Retriever evaluation uses Recall@$^2$9 and MRR@$[i]$0. Task 1 additionally uses NDCG@$[i]$1 and Hit@$[i]$2. Task 2 introduces Position-Aware Citation Accuracy (PACA@$[i]$3), which penalizes late correct answers. The benchmark also reports Citation Diversity Entropy and hallucination rate, making it possible to study diversity and verifiability alongside retrieval accuracy (Zheng et al., 21 Jan 2026).

5. Empirical results and observed trade-offs

On ALCE across ASQA, ELI5, and QAMPARI, C$[i]$4-Cite++ is reported to outperform the SOTA baseline by an average of 5.8% in citation quality and 17.4% in response correctness. The detailed summary further states that it has higher throughput than baselines and that removing contextual-aware embedding causes substantial drops: 25.6% in citation F1 and 21.7% in correctness. The ablation is used to argue that semantically encoded citation tokens are central rather than incidental (Yu et al., 19 Nov 2025).

The cited-text-span line reports that CTS-based generation is more factual than abstract-only conditioning. In the summarized Table 4, Abstract input yields ROUGE-L $[i]$5, QuestEval $[i]$6, and ANLI $[i]$7, whereas Oracle CTS yields ROUGE-L $[i]$8, QuestEval $[i]$9, and ANLI $\langle c_i \rangle$0. Keyword CTS yields ROUGE-L $\langle c_i \rangle$1, QuestEval $\langle c_i \rangle$2, and ANLI $\langle c_i \rangle$3. For generation quality, the summarized Table 5 reports ROUGE-F1 $\langle c_i \rangle$4 for Abstract, $\langle c_i \rangle$5 for Oracle CTS, and $\langle c_i \rangle$6 for Keyword CTS, with corresponding BLEU and METEOR gains. The paper also notes that distantly labeled CTS can match or exceed human-labeled CTS for downstream generation (Li et al., 2023).

KG-CTG reports automatic gains from augmenting prompts with knowledge-graph triples. Without knowledge graphs, Vicuna achieves 14.15 METEOR, 12.88 ROUGE-1, 1.52 ROUGE-2, and 10.94 ROUGE-L. With knowledge graphs, Alpaca improves by 36.98% in ROUGE-1 and 33.14% in METEOR, while Vicuna remains competitive with 13.80 METEOR, 12.87 ROUGE-1, 1.48 ROUGE-2, and 10.96 ROUGE-L (Anand et al., 2024).

CiteRAG frames retrieval as the main lever for citation prediction quality. The benchmark reports that RAG universally improves citation prediction by 40–170% on Recall@20, NDCG, and PACA@20 over closed LLMs in zero-shot settings; that CitationRetriever-8B nearly doubles retrieval MRR compared with general embedding models; that multi-level fusion outperforms any single-level retrieval; that CitationGenerator-30B with RAG reaches Recall@20 $\langle c_i \rangle$7 and NDCG@20 $\langle c_i \rangle$8 on Task 1 and PACA@20 $\langle c_i \rangle$9 on Task 2; and that RAG reduces hallucination from 17% to 5% for CitationGenerator-30B. The summary also states that deep retrieval with $$c_i = \mathcal{F}_{\text{contextual}}(e_1^{(i)}, e_2^{(i)}, \dots, e_N^{(i)}).$$0 is often optimal (Zheng et al., 21 Jan 2026).

A holistic attribution study, however, emphasizes that G-Cite should be understood relative to P-Cite rather than in isolation. Its abstract reports a consistent trade-off between coverage and citation correctness, identifies retrieval as the main driver of attribution quality in both paradigms, and recommends a retrieval-centric, P-Cite-first approach for high-stakes applications, reserving G-Cite for precision-critical settings such as strict claim verification (Saxena et al., 25 Sep 2025). At the same time, the detailed summary reports specific G-Cite strengths, including moderate latency and a human evaluation in which G-Cite outputs were judged more correct (78%) than P-Cite outputs (69%), with slightly lower citation hallucination rates (37% vs. 41%) (Saxena et al., 25 Sep 2025).

6. Adjacent problems: recommendation, auditing, and answer-level absorption

One adjacent problem is citation recommendation at writing time. MasterSet defines must-cite citation recommendation for AI/ML literature, focusing on citations whose omission could misrepresent novelty or undermine reproducibility. The benchmark contains 153,373 papers from 15 venues and 2,005,387 citation instances, and labels citations under three schemes: experimental baseline status, core relevance on a 1–5 scale, and intra-paper mention frequency. The task is deliberately framed to simulate generation-time citation auditing from only title and abstract. The reported results are difficult: no method exceeds 50% Recall@100, with SciBERT-NTX reaching 0.383, 0.410, and 0.493 for the three label types (Ratul et al., 20 Apr 2026).

A temporally aware variant of citation recommendation models time preference as a probability distribution over publication time slices and uses that signal to re-rank content-based candidates. The re-ranking step multiplies initial similarity by a sigmoid-transformed probability for the candidate’s time slice:

$$c_i = \mathcal{F}_{\text{contextual}}(e_1^{(i)}, e_2^{(i)}, \dots, e_N^{(i)}).$$1

On PubMed and DBLP, the method yields small but consistent gains over content-based filtering and is presented as a flexible add-on to other citation recommendation frameworks (Ma et al., 2021).

Another adjacent measurement problem is whether a citation is merely selected or actually used. A two-stage framework for Generative Engine Optimization distinguishes citation selection, measured by the number of valid citations per prompt, from citation absorption, measured by an influence score combining repeated references, early appearance, paragraph coverage, TF-IDF cosine similarity, and n-gram overlap. On 602 prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity, the study reports 21,143 valid search-layer citations, 23,745 citation-level feature records, 18,151 successfully fetched pages, and 72 extracted features. Citation breadth and citation depth diverge: ChatGPT averages 6.88 citations per prompt with mean influence 0.2713, Google 12.06 with 0.0584, and Perplexity 16.35 with 0.0646. High-influence pages are reported to be longer, more structured, semantically aligned, and richer in extractable evidence such as definitions, numerical facts, comparisons, and procedural steps (Kai et al., 28 Apr 2026).

This broader measurement perspective matters for G-Cite. A plausible implication is that evaluation based only on whether citations appear is insufficient. The cited literature increasingly separates citation presence, citation correctness, retrieval adequacy, must-cite coverage, and answer-level absorption, indicating that generation-time citation is not a single task but a family of tightly coupled retrieval, alignment, generation, and verification problems.