ReCite: Agentic Reasoning for Faithful Citation
Yuyang Huang, Bobo Li, Jiajia Song, Yuzhe Ding, Chong Teng, Fei Li, Donghong Ji
cs.CL
Quick verdict
A 4B agentic framework beats far larger LLMs on faithful citation (39.15 strict F1); strong, well-documented work, but the 200-paragraph test set tempers it.
Summary
ReCite is a decoupled agentic framework for automatic citation recommendation that targets misattribution—citing real papers that do not logically support a claim—as distinct from fabrication. Built on Qwen3-4B, it chains three modules: CiteLocator (marks citation positions and Mandatory/Optional necessity via sequence restoration with an asymmetric loss), QueryPlanner (SFT plus GRPO to generate intent-aware retrieval keywords), and a Master Brain that queries the Semantic Scholar API, verifies claim-evidence consistency, and retries with refined keywords on failure. Training data comes from 10,893 arXiv LaTeX packages (2024–2025, 93,403 citations): DeepSeek-V3 retroactively synthesizes reasoning traces, necessity labels, keywords, and 2,000 observe-think-act trajectories; a five-LLM consistency check reports 90.6% accuracy (κ=0.805). On a 200-paragraph, 1,023-paper test set, ReCite-SFT(CAP-8) reaches 39.15 strict F1, beating DeepSeek-V4-Flash (33.10) and other far larger baselines, with Lenient (LLM-judged) F1 of 55.14 and Position-Only F1 of 89.71. CiteLocator reaches 66.37 overall F1 (next best 48.04) and narrows the gap to human experts on the out-of-domain BUSCALDI benchmark (75.5 coarse / 52.8 fine-grained F1 vs 81.8/61.7 human). Ablations show large drops without CiteLocator (2.75 strict F1) or QueryPlanner (15.77), while removing the reflective loop costs only 1.14 strict points. Limitations: text-only, CS-only training, no significance testing, an unspecified LLM judge, and no code-release statement.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Misattribution, not fabrication, is the core open citation-hallucination problem in the LLM era, and citation should be treated as claim-level reasoning rather than similarity search (§1).
- A 4B decoupled agent (CiteLocator + QueryPlanner + Master Brain) reaches 39.15% strict end-to-end F1, outperforming far larger generative models including DeepSeek-V4-Flash (33.10%) and DeepSeek-V4-Pro (31.77%) (Table 3).
- Every module is necessary: removing CiteLocator or QueryPlanner collapses strict F1 to 2.75% or 15.77% (Table 7).
- Fine-grained CAP-8 intent supervision improves both strict F1 (39.15 vs 37.47) and Lenient F1 (55.14 vs 47.10) over plain SFT (Table 3).
- CiteLocator's rhetorical signals generalize out-of-domain, beating GM-s2orc-H on fine-grained BUSCALDI placement (52.8 vs 50.0 F1) though still trailing human experts (61.7) (Table 6).
Strengths
Clean problem decomposition with metrics that map onto it
§1 splits citation hallucination into fabrication (claimed largely solved by retrieval) vs misattribution (real but logically unsupported papers), and the three-level evaluation (Overall-Strictly / Lenient / Position-Only, Appendix B.1) directly operationalizes that split.
Verifiable large-scale supervision from LaTeX ground truth
§3.1: 10,893 arXiv LaTeX packages (ACL/EMNLP/CVPR/ICCV/NeurIPS 2024–25) preserve \cite commands and bibkeys, yielding 31,118 location samples, 7,079 planning samples, and 2,000 trajectories; an explicit audit confirms none of the 2,640 training citation decisions overlap the 1,023 test papers (§3.4).
Quality-controlled synthetic data
Evaluation breadth unusual for system papers
Strong efficiency result, honestly framed
Weaknesses
Headline claims rest on a 200-paragraph test set with no statistical testing
Lenient metric depends on an unspecified LLM judge
Baseline fairness concerns
The reflective loop is the weakest link in its own ablation
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper reframes automatic citation recommendation (ACR) as claim-level reasoning rather than similarity search, splitting citation hallucination into *fabrication* (non-existent papers, claimed largely solved by retrieval architectures) and *misattribution* (real papers that fail to logically support a claim). ReCite is a decoupled pipeline of three Qwen3-4B modules — CiteLocator, QueryPlanner, and Master Brain — trained on trajectories synthesized by DeepSeek-V3 from 10,893 arXiv LaTeX packages (2024–2025, 93,403 citations). Central claims: (1) misattribution is the core open problem; (2) the 4B agent beats far larger generative models at strict citation accuracy (39.15% strict F1 vs DeepSeek-V4-Flash 33.10%, Table 3); (3) all three modules are indispensable (Table 7). A project page is linked, but the text contains no code or data release statement.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does ReCite solve?
It targets misattribution in automatic citation recommendation: systems that retrieve real papers which do not logically support the claim being written. This is treated as distinct from fabricating non-existent papers, which retrieval-augmented architectures already largely prevent because retrieval only returns authentic database entries.
How well does ReCite perform compared to large LLMs?
On a 200-paragraph test set with 1,023 target papers, ReCite-SFT(CAP-8), built on a 4B model, reaches 39.15% strict end-to-end F1 (exact ground-truth paper match) and 55.14% lenient F1 scored by an LLM judge, beating much larger models such as DeepSeek-V4-Flash (33.10% strict) and DeepSeek-V4-Pro (31.77% strict).
Is code or data available for ReCite?
The paper links a project page (hyy279.github.io/ReCite) and provides detailed prompts, hyperparameters, and dataset statistics in its appendices, but the reviewed text contains no explicit statement that code or the synthesized trajectory dataset will be released; retrieval also depends on the live Semantic Scholar API.
What are the main limitations of the paper?
The system is text-only, trained exclusively on 2024–2025 computer-science papers; end-to-end conclusions come from a single 200-paragraph test set with no significance testing; the lenient metric relies on an unspecified LLM judge; and the paper's own ablation shows the reflective loop adds only about 1 strict F1 point.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-embedded text (Figures 5–7 axes/labels) is garbled or interleaved by PDF extraction; figure-level findings were read from captions and the surrounding prose, which quote the key values (e.g., the k=8 optimum).
- Institutional authors render oddly in the reference list ('GLM et al., 2024', 'Kimi et al., 2026'); baseline identities were cross-checked against Appendix A before use.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.