Post-hoc Citation (P-Cite) Methods

• P-Cite is a two-stage paradigm that decouples content generation from citation attachment, ensuring verifiable and comprehensive responses.
• It employs retrieval-augmented methods and iterative regeneration to match claims with external evidence, balancing coverage and correctness.
• Benchmark evaluations show that P-Cite systems achieve higher citation coverage and reliable attribution in high-stakes applications compared to G-Cite methods.

Post-hoc Citation (P-Cite) is a two-stage paradigm for source attribution in LLM systems in which the model first produces a draft answer without in-text citations and a secondary mechanism then attaches, adds, or verifies citations against external evidence. In comparative evaluations against Generation-Time Citation (G-Cite), which produces answers and citations in one pass, P-Cite methods achieve high coverage with competitive correctness and moderate latency, while G-Cite methods prioritize precision at the cost of coverage and speed (Saxena et al., 25 Sep 2025).

1. Definition and conceptual scope

In the P-Cite paradigm, citation selection is decoupled from response generation. The first stage produces a citation-free draft; the second stage processes the completed draft and assigns citations, typically by matching generated text to retrieved documents, sentences, or other evidence units. This decoupling permits attribution decisions to be made with access to the entire generated response and the retrieved evidence pool, rather than token-by-token during decoding (Saxena et al., 25 Sep 2025).

This design gives P-Cite a distinctive operational objective. Rather than constraining the decoder to emit citations jointly with content, it treats attribution as a post-generation inference problem over the full answer. The stated advantages are a global evidence context and the ability to support comprehensive attributions across long, synthesized responses. The same framing also supports citation verification, source correction, and regeneration of unsupported content, rather than only source insertion.

P-Cite is primarily discussed in information-seeking and high-stakes settings such as healthcare, law, academia, and finance, where trustworthy outputs must cite human-verifiable sources. Within that setting, the core question is not merely whether a response contains citations, but whether the citations are correct, sufficiently comprehensive, and aligned with the claims made in the generated text (Saxena et al., 25 Sep 2025).

2. Core workflow and technical architecture

A typical P-Cite pipeline has three steps: draft generation, retrieval, and citation attachment or verification. The retrieval stage is optional in the abstract formulation, but empirical results identify retrieval as the main driver of attribution quality. The attribution stage can rely on rule-based matching, retrieval-augmented scoring, reranking, or entailment-style verification modules (Saxena et al., 25 Sep 2025).

A representative implementation is "Citation-Enhanced Generation for LLM-based Chatbots" (Li et al., 2024). CEG decomposes a response $R$ into claims $R = \{R_1, R_2, \dots, R_n\}$, retrieves candidate evidence for each claim, applies a natural language inference (NLI)-based citation generation module, and regenerates the response if any claim remains unsupported. In that system, retrieval is performed from Wikipedia passages segmented into 100-word documents, and ranking is based on cosine similarity: $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$ where $e(\cdot)$ denotes the encoder representation of a claim or document. Sentence boundaries are detected using NLTK, dense retrieval uses SimCSE-BERT, and the NLI module can be instantiated with GPT-3.5 or GPT-4.

The regeneration stage is central to the strongest P-Cite systems. In CEG, if any claim is labeled nonfactual, the system constructs a new prompt containing the original query, the unsupported claims, and the retrieved evidence, then requests a corrected response. This cycle repeats until all statements are supported or a retry limit is reached. The paper characterizes CEG as a training-free plug-and-play plugin that can be attached to various LLMs (Li et al., 2024).

3. Method families and representative systems

Empirical work groups P-Cite methods into zero-shot prompting, retrieval-augmented generation pipelines, entailment-based citation modules, and specialized post-processing or fine-tuned citation recommenders. The unifying feature is that the attribution decision is made after an answer draft exists; what varies is the mechanism by which support is detected or imposed (Saxena et al., 25 Sep 2025).

Method family	Mechanism	Representative notes
Zero-shot P-Cite	Prompt-based post-hoc citation insertion	High precision in some settings, weak recall in others
RAG-style P-Cite	Retrieve evidence, then attach citations	Retrieval quality is the main lever
CEG	Iterative retrieval, NLI citation judgment, regeneration	Training-free and plug-and-play
CiteBART	Fine-tuned encoder-decoder local citation recommendation	Trained on scientific datasets
CiteFix	Post-processes RAG outputs to check or correct cited sources	Uses keyword, semantic, BERTScore, or LLM cross-checking

"Generation-Time vs. Post-hoc Citation: A Holistic Evaluation of LLM Attribution" (Saxena et al., 25 Sep 2025) highlights CEG, CiteBART, CiteFix, and generic RAG-style retrieval pipelines as promising P-Cite approaches. CEG iteratively retrieves relevant evidence, attaches citations, and regenerates unsupported statements. CiteBART is described as a fine-tuned encoder-decoder approach for local citation recommendation trained on scientific datasets such as arXiv. CiteFix is a post-processing method that checks and corrects cited sources in RAG outputs using keyword, semantic, BERTScore, and LLM-based cross-checking, with increased citation accuracy and modest latency increases.

A recurring conclusion across these systems is that fine-tuning yields incremental rather than foundational gains, whereas retrieval augmentation is fundamental. This makes corpus construction, retrieval granularity, reranking, and evidence selection central technical variables in P-Cite system design (Saxena et al., 25 Sep 2025).

4. Benchmarking, metrics, and empirical trade-offs

The main comparative benchmark for P-Cite versus G-Cite evaluates four attribution datasets: ALCE, LongBench, REASONS, and FEVER. The reported metrics include citation correctness $F_1$, citation precision, citation recall, coverage, latency, and human ratings. Citation correctness is defined as

$$\text{Citation Correctness (F}_1\text{)} = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}$$

(Saxena et al., 25 Sep 2025).

The aggregate pattern is a coverage-versus-correctness trade-off. On ALCE, CEG reports correctness $0.422$ and coverage $0.748$; on LongBench, correctness $0.115$ and coverage $0.782$; on REASONS, correctness $R = \{R_1, R_2, \dots, R_n\}$0 and coverage $R = \{R_1, R_2, \dots, R_n\}$1; and on FEVER, correctness $R = \{R_1, R_2, \dots, R_n\}$2 and coverage $R = \{R_1, R_2, \dots, R_n\}$3. In the same comparison, G-Cite can achieve higher precision-oriented correctness on FEVER—for example, CoT Citation reaches correctness $R = \{R_1, R_2, \dots, R_n\}$4 with coverage $R = \{R_1, R_2, \dots, R_n\}$5—but this occurs with substantially lower coverage. The paper therefore characterizes P-Cite as the paradigm that more consistently provides broader source attribution, while G-Cite is better aligned with strict precision-critical verification (Saxena et al., 25 Sep 2025).

Latency is not uniformly dominated by either paradigm. Advanced P-Cite methods can be slower because they require draft generation followed by retrieval or verification, but some P-Cite systems remain faster than advanced G-Cite variants. An explicit ALCE example reports P-Cite CEG with correctness $R = \{R_1, R_2, \dots, R_n\}$6, coverage $R = \{R_1, R_2, \dots, R_n\}$7, and latency $R = \{R_1, R_2, \dots, R_n\}$8 seconds, versus G-Cite CoT with correctness $R = \{R_1, R_2, \dots, R_n\}$9, coverage $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$0, and latency $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$1 seconds (Saxena et al., 25 Sep 2025).

Human evaluation reinforces the system-level distinction. Reported answer correctness is $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$2 for P-Cite and $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$3 for G-Cite, while citation hallucination is $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$4 for P-Cite and $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$5 for G-Cite. These results support the claim that post-hoc attribution can yield answers judged as more reliably grounded, even when it does not maximize per-citation precision (Saxena et al., 25 Sep 2025).

5. Principle-driven evaluation of citation quality

A substantial methodological issue in P-Cite research is how citation quality should be evaluated. "CiteEval: Principle-Driven Citation Evaluation for Source Attribution" argues that prevailing NLI-based frameworks, including AIS-style metrics, are a suboptimal proxy because they assess whether a statement is supported by the cited sources while ignoring the full retrieval context, the user query, and the generated response (Xu et al., 2 Jun 2025).

CiteEval formalizes statement-level citation assessment as

$$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$6

where $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$7 is the set of citations for statement $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$8, $$\text{Sim}(R_i, d_j) = \frac{e(R_i) \cdot e(d_j)}{\|e(R_i)\| \cdot \|e(d_j)\|}$$9 is the full retrieved source set, $e(\cdot)$0 is the generated response, and $e(\cdot)$1 is the user query. This formulation broadens evaluation beyond supportiveness. It explicitly addresses missing citations, misleading citations, redundant citations, substandard citations, and statements whose provenance is the query, the response context, or parametric model knowledge rather than retrieved evidence.

The corresponding benchmark, CiteBench, contains 1,200 annotated instances drawn from ASQA, ELI5, MS MARCO, and LFRQA, with top-10 retrieval contexts and sentence-level citations. Annotation proceeds in three steps: context attribution, citation editing, and 1–5 citation rating. Reported inter-annotator agreement is $e(\cdot)$2 for context attribution and $e(\cdot)$3 Krippendorff’s $e(\cdot)$4 for rating (Xu et al., 2 Jun 2025).

CiteEval-Auto provides automated metrics based on model-driven context attribution and citation rating. The paper reports that these metrics achieve substantially higher correlation with human judgment than prior automatic metrics, with Pearson correlation up to $e(\cdot)$5 versus less than $e(\cdot)$6 for NLI-based baselines. For P-Cite research, this is consequential because it shifts evaluation from binary entailment checking toward a more comprehensive assessment of whether the right claims are cited, whether the best available sources were used, and whether some statements should remain uncited (Xu et al., 2 Jun 2025).

6. Relation to broader citation theory and citation relevance

Although P-Cite is an LLM-attribution paradigm, its motivating concerns have antecedents in the quantitative study of human citation behavior. "Theory of citing" analyzes misprints in citations to twelve high-profile papers and finds that the great majority of misprints are identical to earlier misprints, leading to the estimate that about $e(\cdot)$7–$e(\cdot)$8 of scientific citations are copied from reference lists used in other papers. The same work reports that the distribution of misprint repetitions follows a power law and develops a stochastic model in which citation copying helps explain why some papers become highly cited (Simkin et al., 2011).

"Time and Citation Networks" approaches citation networks as temporally ordered directed acyclic graphs and applies transitive reduction to isolate what it calls the causal skeleton of knowledge transfer. For arXiv citation networks, roughly $e(\cdot)$9–$F_1$0 of citation edges are removed by transitive reduction, whereas patent citation networks show much less reduction. The paper interprets the remaining edges as a simple indicator of the important citations and suggests that many references in a bibliography were not directly relevant to the work (Clough et al., 2015).

"CausalCite: A Causal Formulation of Paper Citations" reframes citation-based significance as causal impact on follow-up papers rather than raw citation count. It defines pairwise, total, and average causal impact using text matching over LLM embeddings and counterfactual synthesis, and reports stronger alignment than raw citations with expert judgments and test-of-time awards (Kumar et al., 2023). These studies concern bibliometrics rather than LLM source attribution, but they establish a persistent distinction between citation count, citation presence, and citation relevance. That distinction is directly germane to P-Cite: a system that merely appends sources is not necessarily producing citations that are essential, direct, or causally informative.

7. Limitations, misconceptions, and likely development paths

A common misconception is that P-Cite is only a formatting pass over already accepted content. The stronger P-Cite systems do more than append references: they retrieve evidence, verify support at claim granularity, detect unsupported statements, and regenerate content until support constraints are met. In this sense, P-Cite can function as a post-generation control loop rather than a terminal annotation stage (Li et al., 2024).

Another misconception is that citation evaluation reduces to checking whether cited passages entail the statement. CiteEval shows that this omits the full retrieval context and cannot penalize statements that should have been cited but were not, nor distinguish misleading, redundant, or substandard citations. It also overlooks the fact that some statements are better attributed to the user query, prior response context, or parametric model knowledge, and therefore are not always appropriately judged against retrieval evidence alone (Xu et al., 2 Jun 2025).

Current evidence also argues against treating P-Cite and G-Cite as interchangeable. The reported recommendation is retrieval-centric and P-Cite-first for high-stakes information-seeking applications that require broad source attribution and verifiability across long answers, while G-Cite is reserved for precision-critical settings such as strict claim verification. This division reflects the observed trade-off: P-Cite tends toward high coverage with competitive correctness, whereas G-Cite tends toward higher precision with lower coverage (Saxena et al., 25 Sep 2025).

A plausible implication is that future P-Cite systems will emphasize joint optimization of retrieval, attribution, and revision rather than citation insertion alone. The existing evidence already points in that direction: retrieval quality dominates attribution quality, regeneration improves unsupported drafts, and principle-driven evaluation rewards systems that select the right sources in the broader response context rather than merely attaching any supporting citation.