CiteRAG: Citation-Centric RAG Systems

• CiteRAG is a citation-centric framework that enforces evidence-backed answers through constrained generation and abstention when support is insufficient.
• It integrates specialized retrieval with fine-tuned LLM generation to address both fiscal document intelligence and academic citation prediction.
• The approach emphasizes rigorous verification, fine-grained citation spans, and multi-level retrieval to improve citation fidelity.

CiteRAG denotes a citation-centered orientation within retrieval-augmented generation in which attribution is treated as a primary object of modeling, evaluation, or enforcement rather than as an auxiliary display feature. In current arXiv usage, the label has been applied to at least two distinct artifacts: a “citation-enforced RAG framework for fiscal document intelligence” that prioritises explainability, auditability, and abstention in tax-compliance settings (Shanivendra, 11 Mar 2026), and the “first comprehensive retrieval-augmented generation (RAG)-integrated benchmark for evaluating LLMs on academic citation prediction” (Zheng et al., 21 Jan 2026). Around these two uses lies a broader research program on citation generation, citation correction, evidence verification, and finer-grained attribution in RAG outputs (Qian et al., 2024, Maheshwari et al., 22 Apr 2025, Qian et al., 13 Oct 2025, Chen et al., 25 Sep 2025).

1. Scope and meanings of the term

The recent literature uses “CiteRAG” in more than one sense. In one sense, it names a deployed-style framework for high-stakes document question answering in which the generator is constrained to cite evidence and abstain when support is insufficient (Shanivendra, 11 Mar 2026). In another, it names a benchmark and toolkit for academic citation prediction, with task definitions, a large corpus, specialized retrievers, specialized generators, and reproducible evaluation (Zheng et al., 21 Jan 2026). This dual usage matters because the two artifacts address different technical problems: citation fidelity in answer generation versus citation recommendation over a scholarly corpus.

Usage of “CiteRAG”	Primary objective	Core distinguishing features
Fiscal document intelligence (Shanivendra, 11 Mar 2026)	Cited, explainable knowledge retrieval in tax compliance	Source-first ingestion, page-level provenance, citation-enforced generation, abstention
Academic citation prediction benchmark (Zheng et al., 21 Jan 2026)	Evaluate LLMs on academic citation prediction	Two tasks, 554k-paper corpus, multi-level retrieval, specialized retrievers and generators

The term also sits within a wider family of systems that treat citations as explicit outputs. “Citegeist” applies dynamic RAG on the arXiv Corpus to generate a related work section and other citation-backed outputs, using a mixture of embedding-based similarity matching, summarization, and multi-stage filtering (Beger et al., 29 Mar 2025). CG-RAG, while not named CiteRAG in its title, similarly targets research question answering by integrating sparse and dense retrieval signals inside citation-graph retrieval (Hu et al., 25 Jan 2025). This suggests that CiteRAG is best understood not as a single canonical architecture but as a citation-first RAG design space.

2. CiteRAG as citation-enforced RAG for fiscal document intelligence

In “Citation-Enforced RAG for Fiscal Document Intelligence: Cited, Explainable Knowledge Retrieval in Tax Compliance,” the framework is a “source-first” multimodal retrieval-augmented generation pipeline built for IRS and state tax materials (Shanivendra, 11 Mar 2026). Every PDF is passed through a multimodal extractor that pulls embedded text layers intact, runs OCR over scanned pages, preserves spatial cues for tables and figures by “linearizing” them, and attaches metadata including document ID, page range, and section heading. Segmentation aligns with page breaks, section headings, and minimal overlap windows such as 50 tokens. For the evaluation subset, this produces $N = 10{,}491$ text chunks, each with a unique chunk ID.

Retrieval is defined over chunk embeddings produced by BAAI/bge-large-en-v1.5, normalized to unit length and indexed with exact FAISS. At query time, the question embedding $v_q$ is matched to chunk embeddings $v_c$ by cosine similarity,

$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$

The framework defines retrieval confidence as

$$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$

and reports that $C(q) > 0.55$ empirically correlates with “answerable” queries (Shanivendra, 11 Mar 2026).

Generation is handled by a locally-hosted Llama 3.2 3B-Instruct model conditioned only on retrieved chunks, with no external tool calls or web access. The prompt is strictly “evidence-only”: top-$k$ chunks are shown in full and tagged with $[DocID, page\text{-}range, chunkID]$. Citation enforcement is implemented through a constrained decoding plus post-generation validation loop rather than by retraining the LLM. CitationConstraints require that each paragraph emit at least one citation token sequence, and ValidateCitations checks that each paragraph contains a well-formed citation. If constraints cannot be met, the system returns an abstention response rather than forcing an unsupported answer (Shanivendra, 11 Mar 2026).

Abstention is a first-class behavior. The rule is: if $C(q) < \tau$, return Abstain$(q)$; otherwise continue to generation, with $v_q$0. The abstention response explicitly states that sufficient authority cannot be found in the provided IRS or state documents and reports the top retrieved chunks and why they do not fully resolve the question. The design objective is therefore not merely grounded generation, but conservative, auditable behavior under uncertainty.

3. CiteRAG as a benchmark for academic citation prediction

In “What Should I Cite? A RAG Benchmark for Academic Citation Prediction,” CiteRAG is a benchmark and open-source toolkit rather than a single deployed system (Zheng et al., 21 Jan 2026). It defines two task instances at different granularity levels. Task 1 is coarse-grained list-specific citation prediction: given a query paper represented by title $v_q$1 and abstract $v_q$2, the model outputs a ranked list $v_q$3 of candidate references from corpus $v_q$4. Task 2 is fine-grained position-specific citation prediction: given $v_q$5 plus a local context $v_q$6 around a placeholder $v_q$7 in the body text, the model predicts a single cited paper $v_q$8 from $v_q$9 (Zheng et al., 21 Jan 2026).

The benchmark dataset contains 7,267 instances for Task 1 and 8,541 instances for Task 2. Its corpus contains 554k papers spanning many major subfields, collected through a three-level structure. Level 1 contains domain label, title, and abstract; Level 2 adds the Introduction section; Level 3 contains full text with references stripped plus Conclusion. Preprocessing uses pattern matching for section extraction, removes citation markers to prevent leakage, unifies metadata, and excises papers used for Task 1 and Task 2 from the corpus to avoid test contamination. The remaining approximately 538,911 papers form the retrieval index (Zheng et al., 21 Jan 2026).

The retrieval architecture is a multi-level hybrid RAG pipeline. A fine-tuned embedding model retrieves independently from the three corpora $v_c$0, and the candidate lists are fused by Reciprocal Rank Fusion,

$v_c$1

with $v_c$2 described as a small constant, for example 60 (Zheng et al., 21 Jan 2026). The base embedding model is Qwen-3-Embedding-8B, fine-tuned into CitationRetriever-8B using 30,000 query papers and their cited references to create 400,000 positive pairs, optimized with an InfoNCE objective over cosine similarity and hard negatives.

Generation is handled by CitationGenerator-4B and CitationGenerator-30B, based on Qwen3-4B and Qwen3-Coder-30B, using supervised fine-tuning on Task 1 and Task 2 examples in ChatML format. The prompt contains system instructions, paper metadata, retrieved paper titles and abstracts or longer excerpts, and for Task 2 the local passage containing $v_c$3 placeholders. The generator is trained to emit a JSON list of predicted citations plus optional reasoning (Zheng et al., 21 Jan 2026). In this formulation, CiteRAG is not enforcing citations for factual claims in generated prose; it is directly predicting which scholarly sources should be cited.

4. Core technical motifs across CiteRAG-related research

A central motivation for citation-centric RAG is that retrieval quality alone does not guarantee usable attribution. “On the Capacity of Citation Generation by LLMs” explicitly defines citation generation in RAG as producing an answer $v_c$4, decomposed into statements $v_c$5, with citation sets $v_c$6 such that each cited passage factually supports the corresponding statement (Qian et al., 2024). That paper distinguishes pre-hoc citation generation, where citations are integrated into autoregressive generation, from post-hoc citation generation, where answers are produced first and citations are matched later. It further introduces refined citation metrics intended to avoid over-penalizing unnecessary and excessive citations, and proposes a Generate-then-Refine method in which a refiner LLM revises citation sets without altering the answer text.

A second motif is granularity. “Concise and Sufficient Sub-Sentence Citations for Retrieval-Augmented Generation” observes two concrete problems in existing attribution methods: citations are “typically provided at the sentence or even paragraph level,” and sentence-level citations may omit information essential for verification (Chen et al., 25 Sep 2025). The paper therefore proposes sub-sentence citations that are both concise and sufficient, develops annotation guidelines, constructs a corresponding dataset, and introduces an attribution framework that uses LLMs to automatically generate fine-tuning data while filtering low-quality examples with a credit model. This positions citation span selection itself as a modeling problem rather than as a formatting decision.

A third motif is rigorous verification. VeriCite decomposes attribution into three stages: initial answer generation with NLI-based claim verification, supporting evidence selection through passage utility assessment and evidence extraction, and final answer refinement that merges verified statements and consolidates citation indices (Qian et al., 13 Oct 2025). The design premise is that the system should verify support before synthesis rather than trusting the generator’s in-line citations at face value. CiteFix addresses a similar issue through post-processing: it segments answers into factual points, preserves the original number of citations for each point, reranks the retrieved documents with a similarity function, and replaces citations with the top-ranked supports (Maheshwari et al., 22 Apr 2025). Its proposed methods include keyword plus semantic-context matching, a fine-tuned BERTScore model, and a lightweight LLM-based matcher.

These lines of work clarify the broader technical meaning of CiteRAG. It is not simply RAG plus displayed sources. It is RAG in which citation placement, citation sufficiency, citation precision, evidence verification, and user verification cost become explicit system objectives.

5. Evaluation protocols and empirical findings

The fiscal CiteRAG framework is evaluated on 298 real-world PDF documents—145 IRS, 85 CA-FTB, and 68 NY Tax—split into 10,491 chunks, using 55 manually-constructed queries covering federal and state topics (Shanivendra, 11 Mar 2026). Automatic metrics include retrieval confidence $v_c$7, proportion abstained, and citation format compliance. The framework reports 100% citation format compliance for non-abstained answers. Human evaluation reports Citation Support of 94.5%, Format Compliance of 100%, Abstention Accuracy of 88.0%, Hallucination Rate of 1.8%, and Usefulness of 4.2 on a 1–5 Likert average. The reported interpretation is that nearly all claims in answered outputs were directly supported by cited passages, and abstention was used appropriately when support was low.

The academic citation-prediction CiteRAG benchmark uses both retrieval and generation metrics (Zheng et al., 21 Jan 2026). For retrieval, it reports Recall@$v_c$8 and MRR@$v_c$9. CitationRetriever-8B reaches Recall@20 $$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$0, Recall@50 $$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$1, MRR@20 $$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$2, and MRR@50 $$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$3, exceeding TF-IDF, BM25, All-MPNet-Base-V2, and Qwen-3-Embedding-8B on the benchmark. For Task 1, evaluation includes Recall@$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$4, Hit@$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$5, and NDCG@$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$6; for Task 2, the paper introduces Position-Aware Citation Accuracy@$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$7 (PACA@$$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$8). Citation Diversity Entropy and Hallucination Rate are also reported to characterize output quality beyond simple retrieval accuracy.

End-to-end results on the benchmark indicate that retrieval augmentation materially improves citation prediction. Closed-source LLMs such as GPT-5 improve from Recall@20 of approximately 0.061 without RAG to approximately 0.071 with RAG, and PACA@20 from 0.069 to 0.121. CitationGenerator-30B is reported as weak without SFT or RAG, with Recall@20 approximately 0.010, but reaches Recall@20 $$\mathrm{similarity}(q,c) = \mathrm{sim}(v_q,v_c) = v_q^\top v_c.$$9, Recall@40 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$0, NDCG@20 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$1, and Hit@20 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$2 for Task 1 under SFT plus RAG, while achieving PACA@10 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$3, PACA@20 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$4, and PACA@40 $$C(q) = \max_{c \in \mathrm{corpus}} \mathrm{sim}(v_q,v_c),$$5 for Task 2 (Zheng et al., 21 Jan 2026). The same study reports that RAG improves Citation Diversity Entropy from 2.968 to 3.196 and reduces hallucination from 17.4% to 4.9% for CitationGenerator-30B.

Related attribution work reinforces these patterns. Generate-then-Refine raises average Citation F1 on WebGLM-QA few-shot settings from approximately 68% to approximately 77%, with the largest gains in citation precision (Qian et al., 2024). CiteFix reports a relative improvement of 15.46% in overall citation accuracy metrics, with p90 latency of approximately 0.015 s per factual point for keyword-plus-semantic matching, approximately 0.389 s for fine-tuned BERTScore, and approximately 1.586 s for the lightweight LLM-based method (Maheshwari et al., 22 Apr 2025). VeriCite reports average Citation F1 improvements of 10–15 points versus the strongest baseline across five open-source LLMs, with ablations on ASQA showing that removing NLI verification lowers citation F1 by 8.8 points (Qian et al., 13 Oct 2025). Taken together, these results show that citation quality is empirically separable from answer fluency and often improves most when attribution is given dedicated modeling stages.

6. Limitations, misconceptions, and open directions

A common misconception is that CiteRAG denotes a single standardized method. The current literature does not support that interpretation. One CiteRAG is a citation-enforced answer-generation framework for fiscal documents; another is a benchmark for academic citation prediction with task-specific retrievers and generators (Shanivendra, 11 Mar 2026, Zheng et al., 21 Jan 2026). A plausible implication is that the term presently functions more as a citation-centric research label than as the name of one settled architecture.

The fiscal framework identifies several remaining limitations: OCR noise in low-quality scans can mislead retrieval, purely textual representations may lose subtle visual or layout cues, there is no integrated numeric computation such as tax-calculation engines, and reliance on a single LLM means different models may require prompt or constraint tuning (Shanivendra, 11 Mar 2026). Its suggested extensions include additional U.S. states and international jurisdictions, multilingual support, symbolic reasoning modules for numeric computations, layout-aware multimodal embeddings, and longitudinal studies with professional tax examiners.

The academic benchmark highlights different constraints. Its retrieval-noise study reports that CitationGenerator-30B retains approximately 80% of clean performance at 40% irrelevant context injection, and qualitative examples identify failure modes on “deeply nested or overloaded technical contexts” (Zheng et al., 21 Jan 2026). These observations indicate that context selection remains a bottleneck even when retrieval is strong and the generator is specialized.

Broader citation-RAG work suggests several unresolved issues. Sub-sentence citation generation emphasizes that sentence-level attribution can be simultaneously too broad and insufficiently local (Chen et al., 25 Sep 2025). Generate-then-Refine notes reliance on NLI for both training and evaluation, the cost of a full pass per statement, and the absence of text revision in the refiner (Qian et al., 2024). CiteFix points to attention-map-based attribution as only preliminarily explored and regards the high latency of LLM-based matching as impractical at scale (Maheshwari et al., 22 Apr 2025). VeriCite’s modular pipeline implies that future systems may expose explicit confidence estimates for verification, passage utility, and synthesis stages rather than returning a single undifferentiated answer (Qian et al., 13 Oct 2025).

In that broader sense, CiteRAG names an emerging class of systems and benchmarks in which citation fidelity, provenance exposure, and verification cost are elevated to first-order concerns. Whether instantiated as a conservative, abstaining QA system over regulatory documents or as a benchmark for predicting scholarly references, the unifying principle is that retrieval-augmented generation should not only produce text but also specify, justify, and delimit the evidentiary basis on which that text rests.