When we want to understand a research paper, we typically turn to its abstract. However, an abstract is a static, first-person narrative written before the paper has ever been read by the public. It represents what the authors intended to contribute, which often differs from what the scientific community actually finds useful over time.
A growing body of research suggests that the most accurate representation of a paper’s impact is not found within the paper itself, but in the "citances"—the specific sentences in external papers that reference it. By treating the global citation network as a distributed, peer-reviewed summary, researchers are developing models that capture the collective intelligence of the scientific community.
The Citance as a Micro-Summary
The core mechanism of this research direction is the "citance." As Michihiro Yasunaga and colleagues describe in the CL-SciSumm Shared Task, a citance is a high-signal snippet of text that filters out the jargon and peripheral details of a source document.
Think of a citance as a "micro-summary." While an author might spend pages describing a complex methodology, a citing researcher will often distill that entire section into a single, functional sentence: "We use the XYZ algorithm [1] to handle sparse data." This external perspective provides a concise, descriptive vocabulary that is often more effective for indexing and retrieval than the original author's prose. Cornelia Caragea and Sujatha Das Gollapalli demonstrated this in their work on CiteTextRank, showing that these community-provided descriptions significantly improve keyphrase extraction by focusing on the "essence" of the work as perceived by others.
The Wisdom of the Citation Graph
Relying on a single citation would be risky—one author might misinterpret a paper. However, by aggregating citations from dozens or hundreds of different papers, models can leverage a "wisdom of the crowd" effect.
Vijay Viswanathan and Graham Neubig argue in CitationIE that these external descriptions act as a crowdsourced filter. By looking at how a paper relates to its "neighbors" in the citation graph, models can identify which entities and relationships are truly salient. This shift from internal analysis to graph-based context allows systems to move beyond simple summarization toward impact-based summaries, a concept pioneered by Qiaozhu Mei and ChengXiang Zhai. These summaries reflect the evolving significance of a paper—how its perceived value might shift from its original "Results" to a specific "Method" that later researchers found more foundational.
Grounding and the Problem of Epistemic Drift
A major challenge in using external citations is "epistemic value drift"—the tendency for citing authors to oversimplify, strip away nuance, or even misinterpret the original findings. If a summary is built entirely from noisy citations, it risks propagating these errors.
To solve this, researchers like Arman Cohan and Nazli Goharian propose citation contextualization. Instead of just using the citance text, the model uses the citance as a "query" to find the specific "cited text spans" in the original paper. This grounding process ensures that the summary remains faithful to the source material while still being guided by the community's focus. Chrysoula Zerva showed that summaries generated from these specific cited spans and the abstract are often higher quality and more efficient than those generated from the full, unguided text.
Automating the Literature Review
The ultimate goal of this research is to automate the most tedious parts of scholarship: writing "Related Work" sections and discovering relevant papers.
Traditional models for Related Work generation, such as those developed by Xiaojun Wan, treat the task as a multi-document summarization problem. However, by incorporating the citation graph, newer models like CHANGES by Haopeng Zhang can model the hierarchical discourse of a field.
Tools like CiteSee, developed by Joseph Chee Chang and Andrew Head, bring this into the reader's workflow. By surfacing how a paper was cited in other documents the user has already read, it provides a personalized historical context. This turns the act of reading a single paper into an act of navigating a community's collective memory, significantly increasing the rate at which researchers discover relevant new work.