ReCite: Agentic Reasoning for Faithful Citation

AI summary

Overview

Research area: Natural Language Processing — automatic citation recommendation (ACR), retrieval-augmented generation, and agentic reasoning with reinforcement learning.

Technical level: Intermediate. The building blocks (retrieval, supervised fine-tuning, reinforcement learning) are standard NLP topics, but the three-module agentic architecture and the GRPO reward design require some familiarity with modern LLM training pipelines.

Scope: This paper proposes ReCite, a small (4B parameter) closed-loop agent that replaces similarity-based citation lookup with explicit claim-level reasoning, verification, and self-correction, and demonstrates that it outperforms far larger general-purpose models on strict citation accuracy.

What This Paper Is About

Automatic citation tools built on semantic similarity can find papers that are topically related to a claim but do not actually support it — a failure the authors call misattribution. The paper argues that citing is fundamentally a reasoning task rather than a search task, and builds an agent that locates citation points, infers what each citation is meant to do, retrieves candidates from a real academic database, and verifies that the retrieved evidence logically backs the local claim before accepting it.

Key Contributions

1. A conceptual split of citation hallucination into two failure modes. The authors separate fabrication (inventing non-existent papers, largely solved by retrieval-augmented generation) from misattribution (citing real papers that do not logically support the claim), and position misattribution as the open problem that reframes citation as reasoning rather than retrieval.

2. The ReCite framework. A decoupled agent with three cooperating modules — CiteLocator (citation breakpoint and necessity detection), QueryPlanner (intent inference plus keyword generation), and Master Brain (orchestration, evidence verification, and a reflective re-retrieval loop) — all built on Qwen3-4B and trained with SFT plus Group Relative Policy Optimization.

3. A reasoning-oriented citation dataset. Built from 10,893 LaTeX source packages from 2024–2025 top CS venues, the dataset adds three supervision layers absent from prior resources (RefSeer, S2ORC): location perception with Mandatory/Optional necessity labels, an 8-category citation intent taxonomy (CAP-8), and synthesized multi-turn reflective trajectories including failure-and-recovery paths.

4. Evidence that a small task-specific agent beats massive general models. ReCite-4B exceeds much larger baselines — including a 1.6T-parameter model — on strict end-to-end citation accuracy, with ablations isolating the contribution of each module.

Main Findings

• Strict end-to-end accuracy: ReCite-SFT(CAP-8) reaches 39.15% Strict F1, ahead of DeepSeek-V4-Pro (31.77%), Qwen3.6-Plus (31.34%), and all three general-purpose agents (Claude Code, Hermes Agent, OpenClaw, all at roughly 10–22%).
• Lenient evaluation: The model reaches 55.14% Lenient F1, a gain of about 16 points over Strict, which the authors interpret as evidence that the agent often finds logically valid alternative papers that the original author simply did not cite.
• Position-only accuracy: Near 90% F1, showing that identifying where a citation belongs is close to solved relative to which paper to cite.
• CiteLocator is the strongest single module: 66.37% Overall F1 for citation placement versus 48.04% for the best LLM baseline (Qwen3.5-27B), with gains on both Mandatory and Optional categories.
• QueryPlanner approaches a 671B model: 57.81% reasoning Token F1 after GRPO, close to DeepSeek-V3's 58.63% and well above the 27B open-source baseline, while running at 4B parameters.
• Ablation results are stark and asymmetric: Removing CiteLocator collapses Strict F1 from 37.47% to 2.75%; removing QueryPlanner drops it to 15.77%; removing the Master Brain's reflective verification costs only about 1 point of Strict F1 but 4.34 points of Lenient F1.
• Reflection matters most under uncertainty: The Master Brain's verification loop contributes little when the first retrieval is correct but is the main safeguard against confidently accepting a wrong candidate.
• Retrieval depth saturates: End-to-end F1 improves up to k=8 retrieved candidates and then plateaus, with larger k adding context noise and latency; the authors fix k=8.
• Cross-venue generalization: ReCite's F1 stays balanced across CVPR, EMNLP, ICCV, and NeurIPS test paragraphs, whereas several baselines skew toward particular venues.
• Data validation: A cross-model consistency check on keyword quality reached 90.6% average accuracy (Cohen's κ = 0.805), and an audit confirmed none of the 2,640 training citation decisions touched any of the 1,023 test papers.

Methodology in Plain English

The authors start from raw LaTeX source files rather than PDFs, because LaTeX preserves the exact \cite commands — giving ground-truth citation positions and bibliography keys for free. They strip citations from paragraphs and train a model to reinsert a [#CITE#] marker at each legitimate location, additionally labeling each as Mandatory (specific methods, datasets) or Optional (general background). Because citation markers are extremely sparse relative to ordinary text, they use an asymmetric cross-entropy loss that weights the rare marker tokens more heavily.

For each detected position, a second module infers why a citation is needed. Working backwards from the known target reference, DeepSeek-V3 generates a short reasoning chain and a keyword set that would retrieve that paper. This module is fine-tuned and then further trained with GRPO, where the reward combines three signals: whether the output follows the required format, whether it recalls the correct entities, and whether the inferred intent matches the ground-truth intent.

The third module acts as a controller. It runs the whole workflow as observe-think-act loops: it calls the locator, calls the planner, queries the Semantic Scholar API, and then judges whether any candidate's metadata actually supports the claim. Crucially, if nothing fits, it does not guess — it writes a diagnosis of why retrieval failed, rewrites broader keywords into more specific ones, and tries again. This closed loop is what the authors mean by "reflective," and it is trained on synthesized trajectories that deliberately include failure cases.

The core design decision is architectural: because retrieval always targets a real database, fabricating a paper becomes impossible by construction, which lets the whole system focus on the harder problem of logical support. The evaluation mirrors this three-stage decomposition with three matching strictness levels — exact paper match, LLM-judged relevance, and position-only.

Why This Matters

Impact on research. The paper reframes citation recommendation from a retrieval benchmark into a reasoning and verification problem, and argues that a 4B task-specific agent can beat trillion-parameter generalists at a narrowly defined scientific task. If that holds, it suggests that careful decomposition and task-specific supervision matter more than raw scale for high-stakes factual grounding, and it provides a reusable dataset schema (location, intent, reflection) that other agentic NLP pipelines could adopt.

Real-world applications.

• AI writing assistants used by researchers, which currently risk inserting confident but logically unsupported references.
• Reference managers and bibliographic databases that could offer "does this paper actually support my sentence?" checks rather than keyword search.
• Editorial and peer-review triage tools at publishers, flagging submissions with systematically weak claim-citation alignment.
• Literature review automation, where an agent must assemble the correct set of foundational works for a narrative rather than a ranked list of similar abstracts.
• Privacy-sensitive research environments: the authors emphasize that a 4B model can be quantized and deployed offline, avoiding sending unpublished manuscripts to cloud APIs.

Industry relevance. The work sits at the intersection of two commercial pressures: publishers and academic tooling companies need defensible citation quality, and LLM vendors need to reduce factual hallucination in professional workflows. The offline-deployment argument speaks directly to law firms, pharma, and defense research groups that cannot use hosted models. The result that black-box general agents fail badly here is also a useful signal for anyone assuming that a capable coding or browsing agent will transfer to scholarly verification tasks.

Future Directions

• Multimodal verification. Real citations often depend on figures, charts, tables, or pseudocode, which a text-only verifier cannot assess. The authors propose integrating vision-language models to check multimodal evidence.
• Cross-disciplinary generalization. The entire dataset comes from computer science venues, so it is unknown whether the learned rhetorical signals transfer to the humanities, biomedical sciences, or social sciences, where citation conventions and intent distributions differ.
• Handling subjective author bias and multi-valid citations. The gap between Strict (39.15%) and Lenient (55.14%) accuracy suggests the strict metric partly measures disagreement with the original author's choice rather than system error. Defining and evaluating "valid alternative" citations more rigorously is an open measurement problem.
• Closing the open-ended reasoning gap. The authors concede that a specialized 4B model trails large general models outside the targeted workflow; whether the same decoupled design can be scaled up or combined with a stronger general reasoner remains untested.

Target Audience

This paper is most useful to NLP researchers working on retrieval-augmented generation, agentic frameworks, and factual grounding; to engineers building academic writing tools or reference management systems; to publishers and editorial technology teams concerned with citation integrity; and to machine learning practitioners interested in a concrete case study of how decomposition, synthetic trajectory data, and reinforcement learning can let a small model outperform much larger general-purpose systems on a verifiable, high-stakes task.

Authors’ abstract

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.