Explicit Evidence Grounding via Structured Inline Citation Generation

A new framework called FullCite aims to improve how large language models cite their sources by generating structured inline citations that link each claim to both a source document and a specific piece of supporting evidence, according to research submitted in 2026 [1]. The framework, detailed in a paper submitted to arXiv on 5 June 2026, proposes three distinct strategies for inline citation generation: prompt-based generation, constrained decoding over a citation grammar, and posthoc span alignment [1]. The researchers evaluated FullCite using three question answering benchmarks—ASQA, BioASQ, and ExpertQA—assessing citation quality along three dimensions: document-level correctness, evidence span identification, and claim-citation faithfulness [2]. The evaluation revealed a clear performance gap. While large language models (LLMs) proved generally effective at identifying relevant documents, they struggled to identify the precise supporting spans within those documents [1]. Results showed that document-level identification appears to be the easier sub-task, with Doc-F1 substantially exceeding Snippet-F1 across all settings and datasets, indicating that span localization remains the principal source of difficulty [3]. The posthoc span alignment strategy yielded the largest gains in correct evidence identification. Specifically, it increased snippet-F1 from 12.80 to 61.87 for the ASQA benchmark [4]. The researchers also observed a trade-off between snippet-level localization and claim-citation faithfulness, noting that when claim and citation are generated jointly, it is likely to inflate their semantic similarity independently of whether the citation would substantiate the claim against the source [3]. FullCite's approach contrasts with most previous work by jointly referencing documents and supporting evidence spans, which the authors argue leads to more transparent and faithful grounding than relying on either document references or evidence snippets alone [2]. The paper concludes that achieving faithful attributed question answering will require research to place greater emphasis on precise evidence span identification [1]. This work arrives amid broader efforts to improve citation granularity in language model outputs. A separate line of research has explored subsentence-level fine-grained citations, presenting a dataset called SCiFi with 10,000 Wikipedia paragraphs to evaluate whether models can locate the exact portion of output supported by a cited source [8]. Another recent framework, C2-Cite, introduced a context-aware citation alignment mechanism that encodes citation symbols with contextual information from retrieved documents during fine-tuning, enabling LLMs to treat citation markers as active knowledge pointers rather than passive placeholders [9].