TreeReader: Hierarchical Academic Navigator

• TreeReader is an interactive academic paper reader that reformats documents into a navigable, hierarchical tree with node-level summaries for efficient content assessment.
• It employs a recursive LLM-driven summarization strategy to build higher-level insights from detailed evidence at the leaf nodes.
• User evaluations reveal that TreeReader reduces cognitive load and enhances navigation compared to traditional linear PDF readers, despite some limitations in deep detail extraction.

TreeReader is an interactive academic paper reader that employs LLMs to reformat scientific papers into a hierarchical, navigable structure, with the aim of enhancing user efficiency, reducing cognitive overload, and facilitating granular exploration of scholarly content (Zhang et al., 25 Jul 2025). Rather than presenting documents as linear streams of text, TreeReader decomposes them into a tree reflecting the sections, subsections, paragraphs, and other structural elements intrinsic to academic writing. Each node of this tree is annotated with an LLM-generated summary, enabling users to quickly assess the content and selectively reveal details as needed.

1. System Architecture and Hierarchical Decomposition

TreeReader parses an academic paper’s native structure and converts it into an explicit node-labeled tree where each node corresponds to a logical text unit: section, subsection, paragraph, figure, or table. In the system’s three-column user interface:

• The left column displays the navigable "navigation tree," allowing users to quickly traverse the hierarchy and pinpoint specific topics.
• The central column ("Main contents") presents a concise summary for the currently selected node, with interactive controls for recursive drilling down to child nodes or collapsing up to parent nodes.
• The right column provides context such as original figures, tables, or the full underlying text, granting users immediate access to supporting material for any summary.

This design enables multi-level chunking of information and aligns with cognitive science principles reducing overload by presenting content in manageable, contextually grouped units.

2. LLM–Driven Summarization Strategy

TreeReader’s summarization pipeline is grounded in the use of state-of-the-art LLMs (specifically OpenAI GPT-4o as stated in the paper) for content condensation. Rather than summarizing extensive sections wholesale, the system employs a recursive summarization approach:

1. For each leaf node (e.g., an individual paragraph), an LLM prompt is constructed that outputs a succinct summary—paired, as mandated by the prompt, with explicit “evidence” (the supporting text span).
2. For non-leaf nodes (higher-level sections), their children’s summaries are concatenated and provided as input to the LLM, which then generates a higher-level summary tied directly to underlying evidence.
3. The result is a set of hierarchical summaries: each section’s summary reflects its child nodes’ content, creating a bottom-up, recursive abstraction of the paper.

This approach also addresses LLM input-length constraints and maintains traceability between each summary point and its source text. The process can be formalized as follows: for each section $S$,

• if $S$ is a leaf, compute $s = \text{prompt}(P)$;
• otherwise, collect $\{s_1, s_2, ..., s_n\}$ from each child and compute $s = \text{prompt}(\{s_1, ..., s_n\})$.

The LLM prompt protocol requests summarized “points” together with their “evidence,” returned in a machine-parsable structure (JSON), enabling further transparency.

3. Interactive Exploration and User Workflow

TreeReader enables a mixed reading mode:

• Skimming: The user is presented with top-level summaries for rapid orientation; navigation down the tree reveals increasingly granular content on demand. This allows quick identification of relevant sections without reading superfluous text.
• Deep reading: Users seeking fine details can expand nodes recursively to access paragraph-level summaries and ultimately the original text, supported by direct links to figures and tables for immediate cross-reference.
• The main navigation affords sequential and non-sequential traversal: users can focus on sections of interest and return to parent or sibling sections efficiently.

This interactive, chunked navigation contrasts with the traditional PDF reader’s linear scroll, which often obscures structural context and leads to cognitive overload when searching for specific content.

4. User Study and Quantitative Comparisons

A within-subjects paper was conducted with five graduate-level computer science researchers, each tasked with reading two papers of similar length and complexity: one via TreeReader and one via a standard PDF interface. Two reading paradigms were evaluated:

• Skimming (5-minute time constraint, followed by comprehension questions assessing grasp of high-level concepts).
• Deep reading (up to 25 minutes, followed by open-ended questions requiring detail retrieval).

Outcomes measured included:

• Navigation efficiency—users rated TreeReader higher for ease of finding information and understanding paper organization.
• Cognitive load, as measured by NASA-TLX, was lower for TreeReader than for baseline PDF reading.
• Qualitative feedback highlighted that hierarchical summaries enabled “quick grasping of key ideas” and expedited identification of relevant sections.

Some users found the depth of information in LLM summaries occasionally insufficient for highly detailed deep reading, indicating a trade-off between efficiency and maximum detail retrieval. The paper size limits generalizability but provides evidence for TreeReader’s effectiveness in reducing cognitive load for complex document navigation.

5. Reliability, Traceability, and Limitations

Several design decisions in TreeReader specifically address reliability and verifiability:

• Summary–text linkage: Every summary point refers directly to “evidence” from the source, available for user inspection to guard against LLM hallucination or mischaracterization.
• LLM minimalism: The system delegates only summarization to the LLM, deliberately eschewing open-ended question answering or content synthesis to avoid spurious factual outputs.
• Interaction gaps: TreeReader currently omits features like Ctrl+F or keyword search, which are standard in PDF viewers. Some users reported challenges in extracting very fine-grained details needed for in-depth analysis, in part due to the layer of summarization abstracting away low-level content.

These design choices improve transparency and minimize the risk of misleading users, but future work (as suggested in the paper) may involve integrating keyword search, more dynamic subtree expansions, and enhanced deep reading support.

6. Comparative Position and Broader Context

TreeReader differs fundamentally from typical LLM-powered chatbot or summarization tools by maintaining and utilizing the explicit hierarchical structure of scientific documents. Compared to linear formats such as PDF and HTML—which present content in an uninterrupted stream—TreeReader foregrounds structure, enabling users to identify, access, and verify section-specific concepts without scanning the entire text. This design addresses a gap in current digital document reading tools, where cognitive overload, context loss, and inefficient retrieval of target information remain pervasive.

The system’s architecture and interaction model are motivated directly by an in-depth formative paper of academic reading practices, and the evaluation underscores its value for information-centric tasks common in research.

7. Future Directions

Potential directions for TreeReader enhancement include:

• Enlarging the user paper sample and extending to non-CS domains to validate utility across disciplines.
• Adding capabilities for deep, fine-grained search and more sophisticated content retrieval in deeply nested hierarchies.
• Integrating functionality for user-driven annotations, cross-document linking, or collaborative reading scenarios.
• Further minimizing the risk of LLM-induced errors, possibly by employing hybrid extractive–abstractive strategies or introducing robust comparison mechanisms between summary points and raw text at all hierarchical levels.

TreeReader represents a new paradigm in scholarly document navigation, leveraging advances in LLM summarization while maintaining structural fidelity and user-driven interactivity, and shows promise for large-scale scientific knowledge access and synthesis (Zhang et al., 25 Jul 2025).