TL;DR
LinkNav is an augmented reading interface that automatically surfaces semantic connections between distant, non-adjacent passages within academic papers. By generating contextual questions for each text segment and locating their answers elsewhere in the document, it creates an interactive layer that helps readers navigate complex scientific literature more efficiently.
What Problem Does This Paper Address?
Academic papers are forced into a rigid, linear format, but the scientific ideas they contain are inherently non-linear. Crucial pieces of interconnected information—such as an introductory claim referencing a specific experimental result, or a method detailed later in an appendix—are often separated by pages of text.
Because of this separation, readers frequently miss these non-local semantic connections. The authors highlight that even expert peer reviewers struggle with this: a significant portion of questions raised during peer review are actually already answered elsewhere within the submitted manuscript. LinkNav aims to bridge these distant passages automatically, reducing the cognitive load required to synthesize complex papers.
Methods
LinkNav establishes intra-document connections using a three-stage pipeline powered by large language models (LLMs):
- Segmentation: The paper is broken down into text segments of approximately 512 tokens, respecting paragraph and section boundaries.
- Question Generation: An open-source LLM (
llama-4-maverick-17B-128E-Instruct) analyzes each segment to generate plausible, insightful follow-up questions that a smart reader might ask. - Answer Detection: The system uses dense embeddings (
text-embedding-small) to calculate similarity between the generated questions and all other segments in the paper. The top 5 most similar segments are passed to an answer-evaluation model (gpt-5-mini), which makes a final decision on whether an answer is present.
If a valid answer is found later in the document, a directed link ($i \rightarrow j$) is created. To prevent cluttering the interface, LinkNav filters out low-utility links, such as questions answered in the same segment, the immediate next segment, or a previous segment.
Datasets
The authors evaluated LinkNav using the PeerQA dataset, which consists of real-world questions from peer reviews in machine learning and natural language processing venues.
For evaluation, they utilized a high-quality subset of 29 papers where authors had explicitly annotated the exact text spans answering the reviewers' questions. They refer to this evaluation subset as PeerQA-Gold.
Key Insights and Ideas
- Reviewers Frequently Miss Existing Answers: An analysis of the PeerQA dataset revealed that 37.6% of all peer-review questions were already addressed in the initial submitted version of the paper. This underscores a clear, practical need for tools that surface hidden information.
- Connections Span Long Distances: On average, the originating question segment and its corresponding answer segment are 10 segments apart. This confirms that LinkNav captures non-trivial, long-range semantic dependencies rather than simple local context.
- Llama-4 is Highly Efficient for Question Generation: When compared to closed-source models like GPT-4o and o3, Llama-4 generated fewer total questions but achieved the highest percentage of answerable questions (~79%). This significantly reduces downstream verification costs and processing time.
Key Results
- Retrieval Efficiency: Limiting the answer search to the top 5 most similar segments successfully captures 72% of the gold-standard answers. This threshold represents an optimal trade-off between computational speed and retrieval recall.
- High-Precision Verification: In the answer detection phase,
gpt-5-miniachieved an F1-score of 0.799 (0.798 precision and 0.801 recall), outperforming GPT-4o's recall (0.586) by a wide margin. High precision ensures that the links surfaced in the user interface are accurate and intuitive.
Limitations and Future Work
- High Generation Waste: Currently, about 65% of all generated questions are filtered out because they link backward in the text, to the same segment, or to the immediate next segment. Future work will focus on optimizing the prompt or model to generate only forward-linking, high-utility questions from the start.
- Evaluation Scale: Due to API costs and time constraints, the full pipeline components were evaluated on a relatively small subset of 29 papers.
- User Interface Preferences: While the system successfully builds the semantic graph, future user studies are required to determine whether readers prefer localized information blocks (like side panels) or direct navigation links to jump through the linear text.
Summary updated