![[2602.18437] FineRef: Fine-Grained Error Reflection and Correction for Long-Form Generation with Citations](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18437] FineRef: Fine-Grained Error Reflection and Correction for Long-Form Generation with Citations
Summary
The paper presents FineRef, a novel framework for improving citation accuracy in long-form generation by addressing citation mismatch and irrelevance through a two-stage training approach.
Why It Matters
As large language models (LLMs) become integral in generating trustworthy content, ensuring citation accuracy is crucial. FineRef's innovative method enhances the reliability of generated information, which is essential for academic and professional contexts where citation integrity is paramount.
Key Takeaways
- FineRef improves citation accuracy by addressing mismatch and irrelevance on a per-citation basis.
- The framework employs a two-stage training strategy combining supervised fine-tuning and reinforcement learning.
- Experiments show FineRef outperforms GPT-4 in citation performance metrics.
- FineRef demonstrates strong generalization and robustness in noisy retrieval scenarios.
- The approach enhances the overall quality of long-form generated content.
Computer Science > Information Retrieval arXiv:2602.18437 (cs) [Submitted on 18 Nov 2025] Title:FineRef: Fine-Grained Error Reflection and Correction for Long-Form Generation with Citations Authors:Yixing Peng, Licheng Zhang, Shancheng Fang, Yi Liu, Peijian Gu, Quan Wang View a PDF of the paper titled FineRef: Fine-Grained Error Reflection and Correction for Long-Form Generation with Citations, by Yixing Peng and 5 other authors View PDF HTML (experimental) Abstract:Generating with citations is crucial for trustworthy Large Language Models (LLMs), yet even advanced LLMs often produce mismatched or irrelevant citations. Existing methods over-optimize citation fidelity while overlooking relevance to the user query, which degrades answer quality and robustness in real-world settings with noisy or irrelevant retrieved content. Moreover, the prevailing single-pass paradigm struggles to deliver optimal answers in long-form generation that requiring multiple citations. To address these limitations, we propose FineRef, a framework based on Fine-grained error Reflection, which explicitly teaches the model to self-identify and correct two key citation errors, mismatch and irrelevance, on a per-citation basis. FineRef follows a two-stage training strategy. The first stage instills an "attempt-reflect-correct" behavioral pattern via supervised fine-tuning, using fine-grained and controllable reflection data constructed by specialized lightweight models. An online self-reflective boots...