citation-aware generation

Existing RAG systems employ coarse-grained attribution (at the sentence or paragraph level), leading to information redundancy or omission of critical evidence, thereby undermining output verifiability. To address this, we propose a clause-level fine-grained attribution framework. First, we establish the first clause-level reference annotation standard and corresponding benchmark dataset. Second, we design an LLM-based automated data generation pipeline incorporating a credibility scoring mechanism to filter high-quality samples, supplemented by human verification. Third, we fine-tune retrieval-augmented generation models to produce clause-level attributions that are precise, concise, and informationally sufficient. Experiments demonstrate substantial improvements in attribution accuracy and user verification efficiency, while preserving readability and informational completeness. Our approach advances RAG interpretability and trustworthy reasoning by enabling granular, evidence-grounded justification.

Large language models (LLMs) in retrieval-augmented generation (RAG) frequently exhibit hallucination, and existing citation generation methods either require extensive annotated data for fine-tuning or struggle to jointly manage multiple evidence sources, resulting in suboptimal performance. Method: We propose VeriCite, the first framework to integrate natural language inference (NLI) models directly into the generation pipeline, establishing a closed-loop “retrieve–verify–generate” paradigm: (1) retrieve candidate evidence; (2) dynamically assess claim veracity via NLI to filter high-confidence evidence; and (3) jointly optimize answer generation and citation selection—without any model fine-tuning. VeriCite incorporates context decomposition, evidence alignment, and utility-aware evaluation. Results: Evaluated across five open-source LLMs and four benchmark datasets, VeriCite significantly improves citation accuracy while preserving answer correctness, consistently outperforming state-of-the-art fine-tuning and post-hoc citation methods.

In high-stakes domains (e.g., healthcare, law, finance), large language models (LLMs) must provide human-verifiable citations to ensure factual reliability and accountability. This work systematically compares two citation paradigms—generation-time citation (G-Cite) and post-hoc citation (P-Cite)—identifying retrieval quality as the primary determinant of attribution accuracy. Within a unified framework spanning zero-shot to retrieval-augmented settings, we conduct multi-scenario experiments across four major attribution benchmarks, complemented by human evaluation and automated analysis. Results show that P-Cite achieves superior trade-offs between citation coverage and correctness, making it better suited for high-risk applications requiring broad, reliable grounding. In contrast, G-Cite attains higher precision but suffers from lower coverage and higher latency, rendering it appropriate for stringent fact-checking tasks where verifiability outweighs breadth. To our knowledge, this is the first empirically grounded, application-aware guideline for selecting citation strategies in LLM-based systems.

Existing RAG systems predominantly employ paragraph-level coarse-grained attribution, which compromises verifiability in long-document question answering. This work introduces ReClaim, the first framework to enable sentence-level fine-grained attribution, achieving per-sentence traceability through alternating generation of claims and their corresponding citations. Methodologically, ReClaim integrates instruction tuning, autoregressive interleaved decoding, and citation-aware decoding control within a RAG architecture to ensure precise, context-aware attribution. On long-document QA benchmarks, it achieves 90% citation accuracy—substantially outperforming state-of-the-art coarse-grained approaches—and demonstrates robustness and reliability in verifiable generation. The core contribution is the establishment of the first end-to-end verifiable generation paradigm that enforces strict claim–citation alignment, thereby advancing trustworthiness and interpretability in RAG-based systems.

This work addresses the challenges of ambiguous citation provenance and content redundancy commonly encountered in existing retrieval-augmented generation (RAG) systems during information integration. The authors propose a knowledge base construction approach grounded in Q&A nuggets, which leverages explicit question-answer semantics to guide information extraction, selection, and generation while preserving source attribution throughout the pipeline. Departing from conventional fuzzy clustering abstractions, the method employs interpretable Q&A fragments as structured intermediate representations, enabling end-to-end traceable reasoning and generation. Experimental results on the TREC NeuCLIR 2024 dataset demonstrate that the proposed approach significantly outperforms the state-of-the-art nugget-based RAG system, Ginger, in terms of nugget recall, density, and citation accuracy.

This work addresses the challenge of automatically generating fact-checking articles grounded in verifiable citations by leveraging claims, veracity labels, and supporting evidence documents. To this end, the authors propose a multi-agent collaborative pipeline that integrates dense retrieval, source-balanced evidence selection, structured content planning, and citation-aware generation. The framework innovatively incorporates a gated self-evaluation mechanism and a natural language inference (NLI)-driven citation auditing module to repair missing citations and automatically eliminate redundant or unsupported references. Experimental results demonstrate that the proposed approach significantly improves citation accuracy and source credibility in the generated articles, thereby validating the effectiveness of jointly optimizing evidence selection, structured generation, and post-hoc citation verification.

Large language models often struggle to precisely cite supporting evidence when generating answers, compromising factual accuracy and traceability. To address this, this work proposes FullCite, a framework that achieves dual alignment of each claim to both its source document and the specific evidential span within it—the first approach to do so. FullCite integrates three strategies: prompt-based generation, citation-aware constrained decoding, and post-hoc span alignment, to produce structured inline citations. Experiments on ASQA, BioASQ, and ExpertQA demonstrate that while large models effectively retrieve relevant documents, they exhibit significant limitations in pinpointing exact supporting evidence spans. FullCite substantially improves output quality across three key dimensions: document relevance, evidence span accuracy, and claim faithfulness.

This work addresses the issue of factual inconsistency between generated answers and cited sources in retrieval-augmented generation (RAG) systems by proposing a corrective RAG pipeline that integrates pre-generation passage filtering with post-generation strict entailment verification. Building upon Corrective RAG and CiteFix mechanisms, the approach further incorporates an LLM-as-judge diagnostic method to enhance citation fidelity and factual grounding. The proposed framework effectively improves the faithfulness of citations while preserving answer relevance and fluency, thereby demonstrating the feasibility of strengthening citation integrity in RAG outputs. Moreover, the study advocates for a new evaluation paradigm that prioritizes strict answer traceability to source evidence, emphasizing the need for more rigorous assessment of attribution accuracy in generative retrieval systems.

This work addresses the lack of verifiable mechanisms in existing retrieval-augmented systems for assessing the true influence of cited sources on generated answers. The authors propose a seven-stage framework that establishes three measurable layers of transparency in multi-hop question answering: answer correctness, citation fidelity, and single-document influence. For the first time, they integrate causal mediation analysis with database provenance theory to formally characterize the “citation-influence gap” and define faithfulness conditions via token-level KL divergence. Leveraging techniques such as citation-aware generation, attribution auditing, and leave-one-out interventions, the system achieves 53.53% answer accuracy and 71.55% average citation fidelity on the HotpotQA validation set, revealing a frequent misalignment between cited references and their actual evidential impact.

This study addresses a critical oversight in current retrieval-augmented generation (RAG) systems: their reliance on human-oriented document representations, which neglect the distinct representational needs of large language models as content consumers. Under fixed retrieval results, the authors systematically evaluate the impact of 14 document representation strategies—including selection, summarization, and rewriting—on question-answering accuracy across four generative models. Introducing answer retention rate as a novel metric to assess whether transformed documents preserve the correct answer, controlled experiments reveal for the first time that answer retention is the primary driver of generation accuracy, challenging prior assumptions that attributed performance gains to specific representational mechanisms. Notably, when answer retention is high, variations in wording, structure, length, or query dependence exert minimal influence on accuracy, underscoring that preserving answer information outweighs representational form.