evidence linking

This study addresses the limitations of prevailing binary support/refutation frameworks in evaluating AI-generated text, which fail to capture the nuanced semantic relationships between generated content and source documents. Moving beyond conventional groundedness paradigms, the work proposes a reader-centered, fine-grained taxonomy of evidential relations by integrating insights from linguistics and philosophy of language, encompassing diverse linkage types such as syntactic rephrasing and inferential strategies. Through theoretical analysis, a human annotation protocol, and benchmark evaluations, the authors systematically demonstrate the feasibility and efficacy of this framework. The resulting approach offers a more transparent and interpretable provenance mechanism for AI outputs, establishing both theoretical foundations and practical pathways for fine-grained evaluation and explainable interfaces in natural language generation systems.

This study addresses the lack of traceability in scientific hypothesis generation. We propose a literature-grounded framework for automated hypothesis generation. Methodologically, we introduce a novel multi-task small language model (SLM) jointly trained on four objectives: reasoning-chain veracity classification, controlled perturbation–robust enhancement, interpretable modeling, and evidence-source alignment; further integrated with retrieval-augmented literature search and structured evidence alignment. Contributions include: (1) pioneering reasoning-chain veracity classification as the primary supervisory signal, improving logical reliability (F1 +22%); (2) achieving an evidence-support score of 0.327 (p<0.01), significantly surpassing the baseline (0.305); and (3) expert evaluations indicating high feasibility and impact (both >3.5/5). Our framework advances hypothesis generation from opaque, “black-box” outputs toward verifiable, high-impact hypotheses supported by auditable, traceable reasoning.

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.

Large language models (LLMs) exhibit opaque internal knowledge utilization, and their generated citations often lack faithfulness and credibility. Method: We propose the context-prior enhanced citation generation task, unifying the modeling of internal knowledge and externally retrieved knowledge in citation synthesis. We introduce RAEL—a novel retrieval-augmented, internal-knowledge-aware, and external-source-verified paradigm—and design INTRALIGN, a method integrating customized data synthesis with knowledge alignment to enable interpretable evaluation across utility, faithfulness, and credibility. Built upon a RAG framework, we establish a multidimensional credibility assessment system. Results: Our approach significantly outperforms mainstream baselines across diverse scenarios. Crucially, we quantitatively uncover the interactive effects of retrieval quality, question type, and model-inherent knowledge on citation credibility—establishing the first diagnostic, quantifiable citation evaluation paradigm for trustworthy AI.

This work addresses the critical challenge of hallucinated citations in scientific texts generated by large language models, which pose a serious threat to academic integrity and are difficult to detect manually. To this end, the authors introduce the first benchmark and verification framework specifically designed for detecting fabricated references in scientific writing. The framework features a novel unified metric for evaluating citation faithfulness and evidence alignment, supported by a large-scale, cross-domain dataset rigorously validated by human annotators. Methodologically, it proposes an interpretable and scalable multi-agent verification pipeline that integrates claim extraction, evidence retrieval, passage matching, and reasoning calibration to enable end-to-end auditing of citation authenticity. Experimental results demonstrate that the proposed approach significantly outperforms existing methods in both accuracy and interpretability, effectively identifying citation errors produced by state-of-the-art large language models and offering a reliable tool for scientific publishing integrity.

Large language models (LLMs) frequently generate hallucinated content and lack verifiable citations when producing fact-intensive text. To address this, we propose a multi-stage self-verification framework that orchestrates a sequential pipeline of *fact verification → reflective revision → citation integration*. The method synergistically combines chain-of-thought (CoT) reasoning with dual knowledge validation—leveraging both internal consistency checks and external authoritative source alignment—to dynamically perform fine-grained factual scrutiny during generation. When inconsistencies are detected, the model triggers reflective revision and automatically annotates traceable, context-aligned citations. Compared to state-of-the-art approaches, our framework substantially reduces hallucination rates while improving factual accuracy and citation reliability. Empirical evaluation demonstrates its effectiveness in high-fidelity applications such as scientific writing and news generation, where trustworthiness and evidential grounding are critical.

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.

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 work addresses the ambiguity in the relationship between cited evidence and supporting claims in long-form question answering, particularly the issue of “evidence boundary overflow.” To resolve this, the authors propose CAGE, a two-stage framework that first constructs an answer-centric cognitive attribution graph to explicitly align semantic units with supporting documents, thereby transforming implicit claim–document matching into structured subgraphs. In the second stage, citation generation is guided by this graph through map-based reasoning. By integrating a plug-in cognitive map induction module with a structured citation inference model, CAGE effectively narrows the attribution space, enhancing both citation faithfulness and claim verifiability. The approach achieves state-of-the-art performance on ASQA, ELI5, and ExpertQA, demonstrating substantial improvements in citation accuracy.

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 challenge of evaluating whether large language models (LLMs) generate scientifically grounded answers with verifiable citation support in reading assistance tasks. To this end, the authors introduce a benchmark dataset comprising 6,211 question–answer pairs spanning eight disciplines and four question types, along with a novel evaluation framework that supports multi-segment valid citations and citation-based refusal mechanisms. A hybrid assessment approach combines a deterministic citation matcher with an LLM-based scorer to systematically evaluate both open- and closed-source models. The results demonstrate that citation-based metrics more effectively discriminate model performance than LLM-as-a-judge scores; notably, open-source models achieve citation accuracy comparable to the best closed-source counterparts while exhibiting 3–6 times lower inference latency.