ReCite: Agentic Reasoning for Faithful Citation

What changed

ReCite introduces a decoupled agentic framework that moves beyond traditional semantic similarity for citation recommendation. The core issue addressed is misattribution, where existing systems cite authentic papers that do not logically support the author's claim, despite advancements in mitigating fabricated references. ReCite argues that accurate citation requires active, claim-level reasoning. The framework orchestrates three key mechanisms: location perception, intent-aware query planning, and reflective verification. It operates in an observe-think-act format, built upon the Qwen3-4B model and reinforced with Group Relative Policy Optimization (GRPO). The agent verifies claim-evidence consistency and employs self-correction loops when retrieved candidates lack logical support. Experiments indicate that this lightweight framework surpasses state-of-the-art massive generative models in strict citation accuracy.

Core Mechanisms:

• Claim-Evidence Verification: Cross-checks candidate paper metadata against the inferred citation intent, accepting citations only when evidence is consistent.
• Intent-Aware Query Planner: Generates retrieval keywords conditioned on predicted citation intents to address context misalignment.
• Reflective Re-retrieval Loop: Analyzes verification failures, rewrites queries, and retries to overcome single points of failure.

Unlike unidirectional pipelines, ReCite's iterative architecture integrates reasoning, retrieval, verification, and reflection. Since retrieval queries target authentic databases, the system eliminates the hallucination of non-existent papers, focusing solely on resolving misattribution. The research also highlights the construction of a large-scale literature dataset to train this agentic workflow, providing supervision for citation location, intent inference, and verification modules.

Why it matters for builders

For AI builders, ReCite presents a significant advancement in developing more reliable AI systems for academic writing and research. The framework's focus on logical reasoning over mere semantic similarity directly addresses a critical limitation in current retrieval-augmented generation (RAG) models. By providing a structured approach to verify claim-evidence consistency and implement self-correction, ReCite offers a blueprint for building agents that can generate more trustworthy and accurate academic content. This is particularly relevant for developers integrating LLMs into research workflows or creating tools that assist in scientific literature review and writing.

Practical impact

Developers can explore implementing ReCite's agentic framework to enhance their citation recommendation tools. The proposed mechanisms for intent-aware query planning and reflective verification can be integrated into existing RAG pipelines to improve citation fidelity. Builders should consider the dataset construction methodology outlined in the paper for training similar agentic citation models. The success of ReCite in outperforming larger generative models in strict citation accuracy suggests that focusing on reasoning and verification can yield substantial improvements in specialized AI applications. Experiment results on a benchmark dataset demonstrate the effectiveness of each component, encouraging builders to adopt or adapt these strategies.

Caveats and source limits

The primary source is a research paper detailing the ReCite framework, published on arXiv. While it reports experimental results demonstrating ReCite's outperformance against state-of-the-art models in strict citation accuracy, independent benchmarks or real-world deployment data are not yet available. The paper mentions the framework is lightweight and built on Qwen3-4B, but specific performance metrics (e.g., latency, resource requirements) are not detailed. The dataset construction is described, but the dataset itself is not publicly released in the provided source. The research is presented as findings from EMNLP 2026, indicating it is a recent academic contribution without information on commercial availability or integration into existing platforms.