Source-linked AI summary
Sensecape: Enabling Multilevel Exploration and Sensemaking with Large Language Models
Sangho Suh, Bryan Min, Srishti Palani, Haijun Xia
TL;DR
Linear conversational interfaces limit complex information work that requires nonlinear exploration, organization, and synthesis. Sensecape integrates LLM-supported exploratory search with canvas and hierarchy views that externalize multilevel abstraction. A user evaluation found that it supports deeper understanding, broader exploration, and more hierarchical knowledge representation, while its learning curve constrains usability for some participants.
Problem
Linear conversational interfaces make complex information tasks difficult because users must gather, organize, synthesize, revisit, and compare information nonlinearly.
Method
Sensecape is an interactive LLM-powered system that combines canvas and hierarchy views to support exploratory search, sensemaking, and navigation across abstraction levels.
Results
Sensecape helped users explore more concepts, develop deeper understanding, revisit information more frequently, and construct hierarchical knowledge representations.
Takeaways & Limitations
Externalizing multilevel abstraction can encourage broader exploration and provide structured support for sensemaking in complex information spaces.
Takeaways & Limitations
The hierarchy view requires users to organize information across multiple abstraction levels, which can be unfamiliar and challenging, while Sensecape’s learning curve led some participants to prefer the baseline.
Abstract
from arXiv · showhide
People are increasingly turning to large language models (LLMs) for complex information tasks like academic research or planning a move to another city. However, while they often require working in a nonlinear manner -- e.g., to arrange information spatially to organize and make sense of it, current interfaces for interacting with LLMs are generally linear to support conversational interaction. To address this limitation and explore how we can support LLM-powered exploration and sensemaking, we developed Sensecape, an interactive system designed to support complex information tasks with an LLM by enabling users to (1) manage the complexity of information through multilevel abstraction and (2) seamlessly switch between foraging and sensemaking. Our within-subject user study reveals that Sensecape empowers users to explore more topics and structure their knowledge hierarchically, thanks to the externalization of levels of abstraction. We contribute implications for LLM-based workflows and interfaces for information tasks.
1 INTRODUCTION
LLMs support information tasks through flexible conversation, but their linear structure limits complex work that requires nonlinear exploration, organization, and synthesis. Sensecape addresses this mismatch by externalizing multilevel abstraction and supporting movement between foraging and sensemaking.
- Problem: Linear LLM conversations can support short questions but make complex information tasks difficult to gather, organize, synthesize, revisit, and compare.Users may lose track of the overall information activity when navigating backward and forward through conversation history.
- Problem: Complex information work alternates between foraging and sensemaking and requires encoding explored information into a representation.The mismatch arises because flexible workflows and organizational approaches do not fit sequential conversation.
- Approach: Sensecape enables exploratory search with LLMs through multilevel exploration and sensemaking across canvas and hierarchy views.Users can switch views to explore and reason at different levels of abstraction.
- Approach: Sensecape externalizes an abstraction hierarchy and enables flexible navigation across levels, providing a more structured representation for information seeking.The system is designed to combine nonlinear exploration with movement between abstraction levels.
- Contributions: The work contributes an interactive LLM interface, externalized multilevel abstraction, and a user study of seamless exploration across semantic levels.The study examines whether this design supports efficient and comprehensive exploration of an information space.
2 RELATED WORK
Prior research characterizes complex information work as exploratory, iterative, nonlinear, and dependent on external representations for sensemaking. Sensecape builds on these ideas by integrating exploration, organization, and multilevel abstraction within an LLM-supported information space.
- Information Exploration and Sensemaking: Exploratory search involves iterative queries, evaluation of diverse sources, and synthesis when people need deeper understanding rather than a specific answer.It is especially relevant when searchers have limited prior knowledge or experience.
- Information Exploration and Sensemaking: Sensemaking encodes information into external representations such as notes, tables, graphs, and concept maps to support synthesis and reflection.External representations reduce the need to retain all information mentally and help reveal connections among facets.
- Information Exploration and Sensemaking: Complex information work is nonlinear and dynamic, requiring shifts between deduction and induction, divergent and convergent thinking, and multiple abstraction levels.Externalizing interconnected facets can itself be cognitively overwhelming and time-consuming.
- Integrating Exploration and Sensemaking: Existing integrated tools still rely on users to articulate information needs as queries and do not guide exploration of knowledge gaps or integration with current knowledge.Sensecape extends this work across multiple abstraction levels in an information space.
- Related Systems: Prior systems support exploration through topic suggestions, faceted search, highlighting, note-taking, clustering, bookmarking, and re-finding information.Large language models additionally synthesize and generate information from large training corpora.
- Visuo-Spatial Organization: Visuospatial organization can reduce cognitive overload, share memory and context, and support manipulation of complex information.Hierarchy views use spatial organization to represent relationships across abstraction levels.
- Visuo-Spatial Organization: Three-dimensional interfaces can be difficult to use because depth and occlusion may create clutter, so Sensecape uses a pseudo-3D hierarchy view with minimal depth and no occluding layers.The design responds to known usability concerns while retaining spatial organization.
- Interaction Techniques: Sensecape implements interaction techniques such as semantic zoom and filtering to support exploration, reasoning, and management of complex information spaces.These techniques build on prior HCI work on controlling information-space complexity.
3 SENSECAPE
Sensecape supports complex information tasks through interconnected canvas and hierarchy views, combining LLM-assisted exploration with spatial organization across abstraction levels. Its features help users generate questions, extract and organize information, manage semantic granularity, and dive into subtopics.
- Core views: Sensecape combines canvas and hierarchy views so users can explore, organize, and reason about information at different semantic levels.The canvas supports searching, gathering, and organizing information, while the hierarchy view shows canvases, topics, relationships, and abstraction levels.
- Canvas view: The canvas provides an infinite whiteboard for adding, grouping, connecting, and directly organizing nodes and generated responses.Users begin interactions by creating a node and can enter topics, statements, or other text as canvas content.
- Canvas view: The expand bar uses node text with Prompt, Explain, Questions, and Subtopics functions to support LLM-assisted exploration.Questions generates 25 questions, while Explain prepends “Tell me about” before sending a prompt to the LLM.
- Hierarchy view: The hierarchy view lets users add higher-level, side-by-side, custom, or LLM-generated subtopic canvases to represent information hierarchically.The motivating workflow illustrates organizing topics such as transportation, rent, and dining under “Cost of Living in San Francisco.”
- Canvas view: Users can extract highlighted response text into nodes and position it for later exploration and organization.Text extraction lets users preserve topics or information from generated responses and place them where desired on the canvas.
- Canvas view: Semantic zoom dynamically changes response granularity, such as showing keywords when users zoom out to identify topics and connections across many nodes.This manages information overload when full response text becomes difficult to read.
- Canvas view: Semantic dive moves a selected node into a lower-level canvas, where users receive recommended subtopics and the hierarchy is updated.Users initiate the transition by double-clicking a node, creating a separate canvas for deeper exploration.
4 USER EVALUATION
The within-subject evaluation compared Sensecape with an integrated conversational-and-canvas baseline across two exploratory knowledge-work tasks, measuring exploration, sensemaking, feature utility, and everyday-work relevance.
- Baseline: The Baseline combined a ChatGPT-like conversational interface with a canvas for creating, grouping, and connecting nodes.Participants could transfer highlighted generated text into the canvas, while Sensecape users additionally had hierarchy-related features.
- Procedure: The study used a within-subject design in which participants explored two topics with Sensecape and Baseline under counterbalanced system and topic orders.Participants imagined preparing a talk and documented their findings while using each system.
- Participants: Twelve participants with varied academic backgrounds and prior experience researching complex topics took part in the 1.5-hour study.Participants were recruited from a local R1 university and mailing list.
- Measures: The evaluation measured prompts, knowledge-structure nodes, concepts, hierarchical levels, revisits to prior information, and perceived feature utility.Concepts were unique domain-specific terms, while sensemaking measures captured hierarchy depth and revisits to earlier topics or representations.
5 RESULTS
The results section analyzes survey responses, think-aloud data, and system logs to assess how Sensecape supported exploration and sensemaking, illustrated by increased concepts, hierarchy, and revisits.
- Overall results: Sensecape participants explored more concepts, structured knowledge more hierarchically, and revisited previously interacted-with information more frequently than Baseline users.These outcomes correspond to the three comparisons summarized in Figure 9.
5.1 RQ1. How does Sensecape support exploration?
Sensecape supported broader exploration without significantly changing prompt, node, or connection counts, suggesting that participants obtained more concepts from a similar amount of prompting.
- Exploration outcomes: 68.3 concepts (SD = 49.1) were explored with Sensecape versus 22.8 (SD = 7.7) with Baseline, a significant difference, t(11) = 3.1, p = 0.01∗∗.Prompt counts, node counts, and connection counts did not significantly differ between systems.
- Prompting: Participants issued similar numbers of prompts with Baseline (M = 5.8, SD = 2) and Sensecape (M = 7.3, SD = 5.2), t(11) = 0.9, p = 0.37.The same passage reports similar node and connection counts across conditions.
- Exploration support: Participants found expand-bar features such as subtopic and question generation helpful for articulating information needs and identifying concepts or terms to explore next.Participants described these features as helping them discover what to look for during exploration.
5.2 RQ2. How does Sensecape support sensemaking?
Sensecape supported deeper hierarchical organization and substantially more revisiting of prior information, while participants also valued its automatic and spatial organization of subtopics.
- Hierarchical structure: Sensecape produced 4.3 hierarchical levels (SD = 1.2) versus 2.6 (SD = 1.6) with Baseline, a significant difference, t(11) = 2.7, p = 0.02∗.Participants described focusing on specific subtopics and diving deeper into them.
- Revisiting information: Participants revisited prior information 12.8 times (SD = 10.9) with Sensecape versus 0.7 (SD = 1) with Baseline, t(11) = 3.8, p = 0.00∗∗.Revisits included previously used prompts or knowledge representations.
- Perceived sensemaking support: Participants said Sensecape helped structure their thinking and reduced the need to organize information manually.Its hierarchy view and automatic organization were described as supporting movement across levels and maintaining structure.
- Spatial organization: Separate canvases for subtopics combined with a hierarchy view helped participants see connections spatially rather than in a purely linear document.Participants contrasted this organization with the potential confusion of a single Miro board or a top-to-bottom Google Doc.
5.3 RQ3. What is the perceived utility of Sensecape’s features?
Participants perceived Sensecape’s features as useful for exploring, organizing, and navigating complex information across multiple levels of abstraction. They especially valued the expand bar, spatial extraction, semantic zoom and dive, and hierarchy view, although the hierarchy view could feel overwhelming.
- Expand Bar: Participants used the expand bar to generate prompts, explanations, questions, and subtopics while exploring information.They averaged 7.25 uses per study, with Prompt used most frequently at 4 times on average.
- Extract and Curate: Extracting response segments and arranging them spatially helped participants organize their notes and thinking.Participants extracted and curated response parts 6.25 times per session on average, then explored those parts 4.2 times per session.
- Semantic Zoom: Semantic zoom helped participants manage information overload by providing different levels of granularity for exploring and organizing complex information.Participants used semantic zoom 12.58 times per session on average, and almost all agreed it supported complex-information work.
- Semantic Dive: Semantic dive supported direct movement between subtopics, helping participants explore and organize complex information while potentially increasing cognitive load and attention demands.Participants used semantic dive 6.58 times per session on average; one participant identified a trade-off between focused exploration and cognitive load.
- Hierarchy View: Ten of 12 participants actively used the hierarchy view to organize information and switch between levels of abstraction.Participants visited it 6.33 times per session on average and reported that it helped them identify connections between topics at different levels.
- Hierarchy View: The hierarchy view distributed information across boards, enabling focused attention on individual topics and deeper exploration of nodes.Participants also reported structuring their thinking and topic knowledge more hierarchically with this view.
5.4 RQ4. How do people see Sensecape being useful in their everyday knowledge work?
Participants envisioned Sensecape supporting everyday knowledge work by helping them explore unfamiliar topics, generate and develop ideas, and share understanding with collaborators. They particularly connected hierarchical organization and visible relationships with complex individual and team tasks.
- To explore and learn about new topics: Participants envisioned Sensecape helping them explore and learn about unfamiliar, complex topics in everyday work.A machine learning engineer said the hierarchical view improved navigation and saved time when learning a new field or technique.
- To generate and develop ideas: Participants saw the hierarchy view as useful for generating ideas, organizing broad topics by subtopic, and finding keywords for further investigation.One researcher described using it to connect seemingly disconnected topics and make arguments easier to reason about than in a single linear block.
- To collaborate: share understanding: Participants envisioned Sensecape helping collaborators share their exploration process and make overall discussion topics easier to understand.One researcher contrasted its visual knowledge structure with the linear documents used in research meetings.
- To collaborate: share understanding: Participants also connected explicit relationships across abstraction layers with onboarding collaborators to complex software systems.A designer described the hierarchy view as useful for organizing information and revealing connections across layers during software-design work.
5.5 Participants Preferred to Use Sensecape for Deeper Understanding of a Topic
Participants generally preferred Sensecape for deeper understanding, while preferences varied by task and user. Sensecape was associated with deeper exploration, whereas the Baseline was viewed as easier to navigate and sometimes better for broad overviews.
- Preference patterns: Seven of 12 participants generally preferred Sensecape, and nine preferred it when seeking a deeper understanding of a topic.All participants favored either Sensecape or the Baseline over linear interfaces without an integrated note-taking area.
- Preference patterns: Participants reported that Sensecape helped them significantly more than the Baseline to dive deeper into subtopics within a topic.The Baseline was reported as marginally easier to navigate, with Wilcoxon signed-rank test p-values reported in Figure 11.
- Context-dependent preferences: Participants described task-dependent preferences: the Baseline supported broad overviews, while Sensecape supported deeper understanding and more complex environments.Interview analysis indicated that having the option to switch between the systems could be advantageous.
- Trade-offs: Sensecape’s advanced features could support more complex tasks and faster work, but participants identified a comparatively steep learning curve.Participants who preferred the Baseline primarily cited the time required to become familiar with Sensecape.
6 DISCUSSION
Sensecape supported broader exploration and deeper, hierarchical understanding, but its flexible interface introduced complexity, novelty, collaboration, and factual-trust challenges. The discussion identifies design directions including adaptive guidance, alternative representations, collaborative exploration, and mechanisms for verifying LLM-generated information.
- Findings: Participants using Sensecape discovered more topics, explored more broadly, and developed deeper hierarchical understanding than with the Baseline.The authors attribute these outcomes to features supporting subtopic, question, explanation, and prompt generation, alongside nonlinear externalization of sensemaking.
- Limitations: Some participants found Sensecape complex and overwhelming despite most organizing and navigating abstraction levels within approximately 20 minutes.The discussion connects this challenge to the interface costs associated with flexible, visually rich interaction.
- Limitations: Novelty effects and limited interaction time may bias perceived usefulness, including whether users recognize the value of the hierarchy view.One participant suggested that longer sessions would reveal the need to organize an increasingly information-dense canvas.
- Limitations: Users with less experience may find nonlinear freedom overwhelming, motivating an option to switch between Sensecape and a locked conversational Baseline.The proposed workflow lets users begin linearly when unfamiliar with a topic and abstract information later.
- Future Work: Sensecape does not support collaborative exploration, although jointly constructing knowledge hierarchies could enable collective information sensemaking.Participants expressed interest in collaborating and sharing their work.
- Future Work: Hierarchies support exploration and sensemaking, but graphs may better expose similarities between information pieces.The authors therefore identify alternative representations as a future direction rather than treating hierarchy as universally optimal.
- Limitations: Real-world deployment requires mechanisms to verify the trustworthiness of LLM-generated factual information because Sensecape intentionally omitted source references during the study.The omission avoided additional search-process distraction and influence on participants’ experience and perceptions.
- Design Implications: Semantic zoom and hierarchy view were appreciated for managing information overload and visual clutter caused by LLMs’ ability to generate large information volumes.The discussion presents these techniques as potentially useful across future LLM information-task systems.
7 CONCLUSION
Sensecape is an LLM-powered system for structured information exploration and sensemaking. The evaluation found broader conceptual exploration, deeper understanding, and more frequent revisiting of information, while multilevel abstraction supported further exploration.
- Conclusion: Sensecape helps users explore more concepts, construct deeper understanding, and revisit information more frequently for a holistic view of complex information spaces.These findings come from the user evaluation study.
- Conclusion: Externalizing multilevel abstraction encourages further exploration and equips users with tools for exploring and making sense of information spaces.The authors frame the work as an initial step toward complex information workflows powered by LLMs.
A APPENDIX
The appendix illustrates how participants used Sensecape’s canvas and hierarchy views to traverse ideas, reorganize exploration, and structure questions and answers across abstraction levels. It also records the prompts used for question, subtopic, and semantic-level features.
- Canvas Examples: P11’s Baseline canvas exploration traversed three connections from global warming’s economic impact to agricultural losses, a “How?” node, and livestock production.The figure documents a sequential path through related concepts.
- Canvas Examples: P12’s Sensecape canvas exploration followed four traversals before the participant repositioned the latest conversation to center the main topic.The reorganization illustrates spatial restructuring after deeper exploration.
- Hierarchy Example: P3’s hierarchy view organized answers about why global warming is a problem into rising sea levels, natural disasters, and disrupted ecosystems across two levels.The hierarchy externalized the participant’s thought process.
- Prompt Details: Table 1 lists prompts for questions and subtopics in Expand Bar and for Lines, Summary, and Keywords in Semantic Zoom.Curly-braced text in the prompt column marks example-input placeholders.