Source-linked AI summary
ReCite: Agentic Reasoning for Faithful Citation
Yuyang Huang, Bobo Li, Jiajia Song, Yuzhe Ding, Chong Teng, Fei Li, Donghong Ji
TL;DR
Automatic citation recommendation must address misattribution because semantically similar papers may not logically support a claim. ReCite uses claim-level verification, intent-aware query planning, and reflective re-retrieval, and it achieves stronger strict citation accuracy than substantially larger generative models. The framework offers a lightweight approach to faithful citation, though its scope remains limited outside targeted text-based computer-science workflows.
Problem
Semantic retrieval can select authentic papers that fail to logically support a claim, making citation misattribution a central challenge for reliable academic writing.
Method
ReCite orchestrates citation location perception, intent-aware query planning, claim-evidence verification, and reflective re-retrieval using synthesized reasoning trajectories.
Results
ReCite achieves stronger strict citation accuracy than substantially larger generative models and general-purpose agents.
Takeaways & Limitations
Grounding literature matching in verifiable logic rather than semantic overlap provides a reliable foundation for automated academic writing.
Takeaways & Limitations
The framework is text-only, trained on computer-science papers, and its cross-disciplinary generalization remains unexplored.
Abstract
from arXiv · showhide
Accurate citations are the foundation of academic writing, tracing intellectual origins and substantiating core claims. However, manually navigating the growing volume of scientific literature is increasingly difficult, prompting reliance on automatic citation recommendation. While modern retrieval-augmented architectures have largely mitigated the fabrication of non-existent papers, current systems relying on semantic similarity struggle with misattribution, often citing authentic papers that fail to logically support the author's claim. To address this challenge, we argue that accurate citation requires a shift from similarity-based search to active, claim-level reasoning. We propose ReCite, a decoupled agentic framework that orchestrates location perception, intent-aware query planning, and reflective verification. Trained on synthesized reasoning trajectories, our agent verifies claim-evidence consistency and triggers self-correction loops when retrieved candidates lack logical support. Experiments demonstrate that our lightweight framework outperforms state-of-the-art massive generative models in strict citation accuracy. By grounding literature matching in verifiable logic rather than semantic overlap, ReCite establishes a reliable foundation for automated academic writing.
1 Introduction
Citation recommendation must move beyond semantic retrieval because authentic papers can still fail to logically support a claim. ReCite addresses this through claim-level verification, intent-aware planning, and reflective re-retrieval, achieving stronger strict citation accuracy than generative baselines.
- Motivation: Semantic relevance does not guarantee logical support, making misattribution the central citation challenge after fabrication is largely addressed.The paper distinguishes invented references from authentic papers that fail to support the cited claim.
- Motivation: Different citation intents require distinct verification criteria, while unidirectional pipelines cannot recover from faulty retrieval.The paper identifies background introduction and method comparison as examples of differing citation contexts.
- ReCite: ReCite reframes citation as reasoning by combining claim-evidence verification, intent-aware query planning, and reflective re-retrieval.Failed verification triggers cause analysis, query refinement, and another retrieval attempt.
- ReCite: ReCite is trained through supervised finetuning and reinforcement learning within an agentic framework for faithful citation.Its workflow follows an iterative reasoning, retrieval, verification, and reflection architecture.
- Results: ReCite substantially outperforms strong generative baselines in strict citation accuracy, while experiments expose citation reasoning as challenging for current LLMs.The reported gains are attributed to the framework's integrated reasoning mechanisms.
- Contributions: The paper separates citation hallucinations into fabrication and misattribution, establishing misattribution as the core LLM-era challenge.This reframing motivates treating citation as a reasoning task rather than a retrieval task.
2 Related Work
Prior citation recommendation systems have shifted toward contextual and retrieval-augmented generation, but semantic similarity can bypass strict logical entailment. Verification and agentic reasoning methods motivate workflows that detect, verify, and retry citation retrieval.
- Citation Recommendation: Citation recommendation has progressed from global and two-stage pipelines toward localized generation and dynamic retrieval-augmented frameworks.These approaches improve contextual integration but remain related to the broader citation-fidelity problem.
- Citation Recommendation: Semantic similarity as a generative prior can bypass strict logical entailment and reduce citation fidelity.The related-work discussion identifies this as a limitation of contextual generation approaches.
- Logical Verification and Agentic Reasoning: Static post-hoc verification cannot correct errors propagated through a unidirectional generation pipeline.Agentic frameworks and step-level reinforcement introduce self-correction as an alternative direction.
- Dataset: Figure 2 reports the distributions of academic conference sources and citation-intent categories in the utilized citation dataset.The caption identifies the two distribution dimensions but does not state a specific comparison.
3 Dataset Construction
ReCite's dataset combines citation-location, intent-aware query-planning, and reflective trajectory supervision from recent scientific papers. Unlike static context-to-reference datasets, it captures the decisions and recovery steps needed for an agentic citation workflow.
- Dataset design: The dataset trains three modules to detect citation positions, infer citation intent and retrieval queries, and verify or retry candidate evidence.These modules are CiteLocator, QueryPlanner, and Master Brain.
- Data source and preparation: 10,893 LaTeX source packages from 2024–2025 computer-science venues preserve explicit citation commands and bibliography keys for supervision.The corpus includes ACL, EMNLP, CVPR, ICCV, and NeurIPS papers.
- Data source and preparation: The preparation pipeline extracts introductions, normalizes citation macros, removes non-textual noise, and filters paragraphs with abnormal citation density.
- Dataset design: Compared with RefSeer and S2ORC, the dataset additionally captures citation location, intent, and reflective trajectories rather than only static context-to-reference mappings.
- Query planning data: The QueryPlanner subset contains 7,079 high-density paragraphs and generates reasoning chains plus keyword sets for retrieving target references.
- Reflective trajectory synthesis: The reflective subset contains 2,000 observe-think-act trajectories, including gold paths and reflection-enhanced variants for retrieval-failure recovery.Reflection-enhanced variants require diagnosing drift, refining keywords, and issuing secondary retrieval.
- Evaluation split: The end-to-end test set has 200 disjoint paragraphs and 1,023 unique target papers, with an audit finding no overlap between training citation decisions and test papers.
4 Method
ReCite decouples automatic citation into location perception, intent-aware query planning, and reflective verification orchestrated by a central agent. Its modules use targeted retrieval and self-correction when candidates fail to support the local claim.
- CiteLocator: CiteLocator identifies citation boundaries and labels their necessity as Mandatory or Optional.
- Framework overview: ReCite separates citation processing into CiteLocator, QueryPlanner, and Master Brain stages that operate iteratively.
- QueryPlanner: QueryPlanner connects local context with retrieval metadata by producing a reasoning chain and keyword set for each citation position.It is fine-tuned with supervised learning and optimized with GRPO using groups of four candidate outputs.
- Master Brain: Master Brain locates citation positions, invokes QueryPlanner, retrieves candidate metadata from Semantic Scholar, and verifies claim-evidence consistency.
- Reflective verification: When retrieved candidates lack consistent evidence, Master Brain attributes the failure, refines keywords from broad domains to specific entities, and performs secondary retrieval.The framework targets misattribution while authentic database retrieval prevents fabrication by construction.
5 Experiments
Experiments evaluate ReCite end-to-end, at the module level, across citation intents and retrieval settings. ReCite achieves strong strict citation accuracy, with its decoupled modules and reflective verification contributing to performance.
- Evaluation setup: The evaluation reports Precision, Recall, and F1 across Strict End-to-End, Lenient, and Position-Only metrics, with CAP-8 denoting citation-taxonomy fine-tuning.
- End-to-end results: ReCite-SFT(CAP-8) achieves the highest Strict F1 of 39.15%, outperforming DeepSeek-V4-Pro, Qwen3.6-Plus, and other baselines.
- End-to-end results: The 15.99% gain in Lenient F1 reaches 55.14%, while Strict F1 remains constrained by exact-match evaluation and subjective author biases.
- Module-level evaluation: CiteLocator reaches Overall F1 of 66.37% versus 48.04% for Qwen3.5-27B, with gains across Mandatory and Optional citation categories.
- Module-level evaluation: QueryPlanner-SFT reaches Token F1 of 56.12% versus 58.63% for DeepSeek-V3, and GRPO further improves reasoning and most keyword metrics.
- Ablation study: Removing CiteLocator lowers Strict F1 to 2.75%, removing QueryPlanner lowers it to 15.77%, and removing reflective verification lowers Lenient F1 from 47.10% to 42.76%.
- Robustness and intent analysis: ReCite maintains balanced end-to-end F1 across four conference sub-domains and achieves higher recall than most baselines across CAP-8 intent categories.
- Retrieval candidate analysis: End-to-end F1 improves for ReCite and DeepSeek-V4-Flash through k = 8, after which context noise limits gains and inference cost increases.
6 Conclusion
The paper introduces ReCite as a decoupled agentic framework for faithful citation. Experiments and analyses show stronger strict citation accuracy than larger generative models, while ablations support the contribution of its components.
- ReCite integrates citation location perception, intent-aware query planning, and reflective verification to address citation misattribution.
- ReCite uses retrieval from authentic academic databases to eliminate fabricated references and grounds citation matching in verifiable reasoning.
- Extensive experiments show stronger strict citation accuracy than substantially larger generative models and general-purpose agents.
- Ablation and intent-level analyses confirm that each component contributes to robust citation grounding.
Limitations
The paper identifies limitations in ReCite’s scope and modality. The lightweight model trails massive general-purpose models in open-ended reasoning, while multimodal and cross-disciplinary generalization remain unexplored.
- The lightweight model inherently trails massive general-purpose LLMs in open-ended reasoning outside the targeted citation workflow.
- ReCite currently operates only on unimodal text, although real-world citations may rely on charts or pseudocode.
- Because training used computer science papers, cross-disciplinary generalization to fields such as humanities or biomedicine remains unexplored.
Ethical Considerations
The paper frames faithful citation as an ethical safeguard against fabricated references and related academic misinformation. It also emphasizes privacy and intellectual-property risks from closed-source cloud models.
- Fabricated references can mislead readers and undermine academic integrity, motivating verification against real-world academic databases.
- A lightweight 4B-parameter model is intended to support future offline, on-device citation assistance without compromising sensitive intellectual property.
- Closed-source cloud-based LLMs pose data-leakage and compliance risks when researchers handle confidential or unpublished data.
A Experiment Settings
The experiments compare ReCite with proprietary, open-source, end-to-end, heuristic, and human baselines across citation tasks, including out-of-domain evaluation. Training and evaluation use NVIDIA RTX 4090 D and A800 GPUs with DeepSpeed and vLLM for distributed training and inference.
- Baselines: The evaluation benchmarks ReCite against proprietary and open-source LLMs for end-to-end and module-level performance.The proprietary baselines include GPT-4o-mini, GPT-5.1-Chat, Kimi-k2-preview, GLM, Qwen, Mimo, and MiniMax models; open-source baselines include Qwen and DeepSeek variants.
- Baselines: Three general agents receive one end-to-end citation objective and autonomously orchestrate their tool calls, with all running on Mimo-V2.5-Pro.
- Baselines: Location-only evaluation compares GM-s2orc, GM-s2orc-H, and Scientist human annotators.GM-s2orc is GPT-2-based and fine-tuned on s2orc, while GM-s2orc-H adds post-hoc NLP heuristics.
- Datasets: Out-of-domain generalization is evaluated on 133 sentences sampled from recent arXiv papers and annotated by three senior researchers.
- Implementation: ReCite is trained and evaluated on NVIDIA RTX 4090 D and A800 GPUs, using DeepSpeed and vLLM with PagedAttention.These tools support distributed training and high-throughput inference.
B Details of Evaluation Metrics
The evaluation aligns metrics with ReCite’s three modules: citation location perception, query planning, and candidate verification. It combines strict end-to-end criteria, breakpoint metrics, keyword-quality measures, and synthesized examples of reflective self-correction.
- End-to-end evaluation: Overall-Strictly requires both correct citation-position identification and retrieval of the exact ground-truth paper.
- End-to-end evaluation: Lenient Evaluation accepts highly relevant substitute papers and uses an LLM-as-a-Judge binary score based on local context and candidate metadata.
- End-to-end evaluation: Position-Only evaluates structural perception by checking whether the logical breakpoint is identified regardless of the retrieved candidate.
- Citation location perception: Citation location perception evaluates generated [#CITE#] placeholders against ground-truth breakpoints using Precision, Recall, and F1-score.These metrics are computed separately for Mandatory citations requiring specific evidentiary backing and Optional citations providing general background.
- Intent-aware query planning: QueryPlanner evaluation measures reasoning and keyword overlap, keyword recall, core-entity phrase recall, Jaccard similarity, and LLM-judged keyword quality.Jaccard similarity compares predicted and ground-truth keyword sets, while the judge scores logical soundness and retrieval feasibility from 0 to 10.
- Reflective verification: The synthesized reflective trajectory shows search reformulation after initially broad keywords retrieve irrelevant robotics and driving papers.The agent identifies search drift and replaces general terms with more precise RLHF-related keywords.
D Case Study on the Multi-Option Reality in Citations
The case study shows that strict exact-match scoring can reject an academically appropriate substitute, while Lenient Evaluation recognizes equivalent logical support. This exposes a mismatch between bibliographic identity and citation faithfulness.
- Case study: ReCite retrieves HuMoR for a claim originally supported by GLAMR because both frameworks use gradient-based global optimization for 3D human motion estimation.
- Metric comparison: Strict exact-match assigns HuMoR a score of zero, whereas Lenient Evaluation accepts it as a valid semantic alternative.The example argues that Strict F1 can underrepresent citation faithfulness when multiple papers support a claim.