VArify: A Visual Analytics System for Verifying Knowledge Enhanced Large Language Model Responses in Food Science

arXiv:2606.10177 · cs.HC · Submitted 2026-06-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "VArify: A Visual Analytics System for Verifying Knowledge Enhanced Large Language Model Responses in Food Science".

Jane: Here's the gist of the paper. Graph Retrieval-Augmented Generation (GraphRAG) enables Large Language Models (LLMs) to leverage structured, domain-specific knowledge graph databases for factually grounded responses.

Tom: However, the retrieval of irrelevant or conflicting data can still result in erroneous responses.

Jane: The core problem addressed by the research is that existing verification methods are insufficient for ensuring error-free responses from GraphRAG systems in knowledge-intensive domains.

Tom: The authors conduct a formative pilot study to characterize the challenges of verifying complex, multi-layered data retrieved by GraphRAG systems.

Paper summary: Jane: So to wrap up this VArify paper, they set out to address a problem where Graph Retrieval Augmented Generation models might provide answers that aren't factually grounded because the retrieved data is flawed or irrelevant.

Tom: The thesis is that for knowledge-intensive domains, human verification of the supporting evidence is still necessary, and they conducted a formative pilot study to characterize the challenges of checking complex, multi-layered data retrieved by Graph Retrieval Augmented Generation systems.

Lu: They presented VArify as a visual analytics system that uses a file directory-inspired tree visualization to allow simultaneous exploration of inter-group relationships and intra-group hierarchies within the retrieved evidence.

Meng: The paper is really about making the process of checking these complex retrievals more manageable for people, which is important when you're trying to ship something to production where errors can be costly.

Tom: They showed that their results indicate that this system effectively helps users distinguish between an LLM’s internal parametric knowledge and the actual grounded context they are working with.

Jane: What matters here is how they characterized the challenges, because they found things like anchoring effects and over-trust in structured evidence as major hurdles for verification.

Lu: They also highlighted misunderstandings due to improper visual representations, specifically that participants struggled to trace errors upstream from relationship errors to mapping errors.

Meng: So it's not just about the AI getting the facts wrong; it’s about how the system presents those facts that makes it hard for a human expert to spot where the broken link is.

Tom: The paper also detailed their standard Graph Retrieval Augmented Generation pipeline, which includes prompt rewriting, question expansion, context retrieval through two paths—one for direct LLM calls and one for knowledge graph queries—and finally context summarization.

Jane: It seems like they're laying out the whole process to show exactly where the data flows so that VArify can step in at the verification stage.

Lu: They describe how context retrieval follows two paths, with sub-questions labeled for direct LLM invocations going to a standalone LLM without knowledge graph context, and those labeled for the knowledge graph being answered in multiple steps through queries.

Meng: So they're showing that the AI doesn't just grab one single answer; it does a whole sequence of steps, and VArify needs to visualize all those paths.

Tom: And they show how the final context summarization stage synthesizes a response by combining answers from all those subqueries.

Jane: It’s about making that entire chain visible so users can see exactly what went into the final output and where the potential errors could hide along the way.

Conclusion: Tom: The paper VArify: A Visual Analytics System for Verifying Knowledge Enhanced Large Language Model Responses in Food Science is focused on showing how to ground LLM responses with evidence.

Jane: It’s about moving past just getting an answer and providing a way for domain experts to actively verify that answer by seeing the context clearly.

Lu: The authors propose this visual analytics system as a way to support verification by leveraging file directory-inspired tree visualizations for simultaneous exploration of inter-group relationships and intra-group hierarchies within the retrieved evidence.

Meng: In short, it's about giving experts better tools to navigate the complexity that comes with using these advanced retrieval systems.

Tom: The implication is that this kind of system can empower users to maintain agency and critical oversight when they are interacting with Graph Retrieval Augmented Generation chatbots in safety-critical areas like food science.

Jane: So, it’s about empowering experts to have a clear picture of the data so they don't just blindly accept what the AI spits out.

Lu: The work offers a framework for designing visualizations that support domain-aligned reasoning by visualizing multi-hop relationships through bundled edges between groups.

Meng: It shows how visualization can be used to support verification without causing over-trust, which is a key finding in this study.

Tom: Ultimately, the goal is to show that verification supported by visualizing retrieved context aligns the domain knowledge of expert users with the context retrieved from the knowledge graph.

Jane: So it suggests that we need tools that help people align their expertise with what's being pulled from a database in these AI systems.

University of California, Davis

cs.HC

Submitted: 2026-06-08

Updated: 2026-11-12

Code: https://github.com/linuslyt/VArify

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 76/100

The gist: Graph Retrieval-Augmented Generation (GraphRAG) enables Large Language Models (LLMs) to leverage structured, domain-specific knowledge graph databases for factually grounded responses.

Key concepts

GraphRAG
This is a method where Large Language Models (LLMs) use structured knowledge graphs to find facts. It helps LLMs give answers based on specific, organized domain knowledge rather than just general training data. The system retrieves relevant facts from this graph to ground the LLM's response.
VArify
This is a visual analytics system designed to help users verify GraphRAG outputs. It uses a tree-like visualization that lets users explore complex relationships and hierarchies in the retrieved evidence easily. Its goal is to make it clear how the AI arrived at its answer so experts can check for errors.
Cognitive Forcing Functions (CFFs)
These are design elements in an interface intended to guide user behavior, such as interrupting a habit of trusting the first answer. In VArify, CFFs are used to prompt users to pause and critically evaluate information instead of immediately accepting the LLM's output.
Progressive Disclosure
This is a design technique where information is revealed in stages rather than all at once. VArify uses this by showing a high-level summary first, allowing users to progressively explore deeper details of the graph and evidence only when they need them, preventing cognitive overload.

Terminology

Summary

Graph Retrieval-Augmented Generation (GraphRAG) enables Large Language Models (LLMs) to leverage structured, domain-specific knowledge graph databases for factually grounded responses. p. 0

Where the paper says it

VArify: A Visual Analytics System for Verifying Knowledge Enhanced Large Language Model Responses in Food Science Sam Yu-Te lee , Yan To Linus Lam, Manami Nakagawa, and Kwan-Liu Ma Abstract—Graph Retrieval-Augmented Generation (GraphRAG) enables Large Language Models (LLMs) to leverage structured, domain-specific knowledge graph databases for factually grounded responses. However, the retrieval of irrelevant or conflicting data can still result in erroneous responses.

However, the retrieval of irrelevant or conflicting data can still result in erroneous responses. p. 0
Where the paper says it

VArify: A Visual Analytics System for Verifying Knowledge Enhanced Large Language Model Responses in Food Science Sam Yu-Te lee , Yan To Linus Lam, Manami Nakagawa, and Kwan-Liu Ma Abstract—Graph Retrieval-Augmented Generation (GraphRAG) enables Large Language Models (LLMs) to leverage structured, domain-specific knowledge graph databases for factually grounded responses. However, the retrieval of irrelevant or conflicting data can still result in erroneous responses.

The core problem addressed by the research is that existing verification methods are insufficient for ensuring error-free responses from GraphRAG systems in knowledge-intensive domains. p. 0

Where the paper says it

Although verification support has been identified as one of the key design strategies for avoiding users accepting erroneous responses, existing designs rarely go beyond providing the retrieved evidence or citations that informed the AI’s answer [19]. Simply providing evidence is insufficient for the level of verification required by experts in knowledge-intensive and evidence-focused domains.

The authors conduct a formative pilot study to characterize the challenges of verifying complex, multi-layered data retrieved by GraphRAG systems. p. 0
Where the paper says it

In knowledge-intensive and evidence-focused domains, human verification of the supporting evidence for an LLM response is still necessary. We conducted a formative pilot study to characterize the challenges of verifying complex, multi-layered data retrieved by GraphRAG systems.

Based on these insights, they present VArify, a visual analytics system that leverages a file directory-inspired tree visualization to support simultaneous exploration of inter-group relationships and intra-group hierarchies within the retrieved evidence. p. 0
Where the paper says it

We conducted a formative pilot study to characterize the challenges of verifying complex, multi-layered data retrieved by GraphRAG systems. Based on these insights, we present VArify, a visual analytics system that leverages a file directory-inspired tree visualization to support simultaneous exploration of inter-group relationships and intra-group hierarchies within the retrieved evidence.

How the training signal is built

The authors developed a standard GraphRAG pipeline consisting of four stages: prompt rewriting, question expansion, context retrieval, and context summarization. The prompt rewriting stage rephrases the user’s prompt to better align with the knowledge graph structure. p. 1

Where the paper says it

Prompt rewriting The first stage of the pipeline rephrases the user’s prompt to better align with the knowledge graph. This allows the user to ask questions in natural language that are not specifically phrased to fit the knowledge graph’s structure.

The question expansion module generates multiple sub-questions that could be relevant. p. 1
Where the paper says it

To achieve this, an LLM is provided with the details of the knowledge graph schema and prompted to paraphrase the user’s question in terms of possible nodes and relationships. Question expansion The pipeline then performs question expansion on the rewritten prompt, which generates multiple sub-questions that could be relevant.

The context retrieval stage follows two paths, forwarding sub-questions labeled for direct LLM invocations to a standalone LLM and those labeled for the knowledge graph to be answered in multiple steps. p. 2
Where the paper says it

Context retrieval The execution of sub-questions follows two paths. Sub-questions labeled for direct LLM invocations are forwarded to a standalone LLM, without any context regarding the knowledge graph structure.

Finally, the context summarization stage synthesizes a response by combining answers from all subqueries. p. 2
Where the paper says it

Finally, the query is executed on the knowledge graph database. Context summarization Once all sub-questions have been answered in parallel, their answers are forwarded to the Context Summarization stage.

Where the gains come from

The formative study revealed several key findings regarding user behavior and interface design needs. Key findings include:

Anchoring effect and confirmation bias

Participants found that verification behaviors are strongly influenced by the anchoring effect, where they align LLM responses with entity mapping results rather than critically reflecting on the information.

Over-trust in structured evidence

The presence of structured tabular evidence often increased users’ confidence, leading participants to rely on surface-level matching instead of reconstructing underlying reasoning.

Misunderstandings due to improper visual representations

Participants struggled to trace errors upstream from relationship errors to mapping errors and desired clearer representations of multi-hop relationships.

Pruning redundant reasoning tokens

Based on these findings, the authors derived three design requirements (DRs) for supporting verification in the food science domain. These include:

DR1: Mitigate cognitive biases through interface design

This requires incorporating cognitive forcing functions (CFFs) such as delaying answer reveal to interrupt habitual reliance on the first response.

DR2: Support structured and domain-aligned reasoning

This necessitates bespoke visualizations to accurately visualize multi-hop relationships and hierarchical groupings among entities.

DR3: Support verification of alignment between LLM responses and evidence

The system must support progressive disclosure and effortless navigation to trace relationships between coordinated views to determine what knowledge is relevant or trustworthy.

VArify was developed with two main components: a chat panel on the left and a graph panel on the right. The chat panel incorporates a cognitive forcing function (CFF) by first showing an Answer Context Summary panel before revealing the final answer. p. 4

Where the paper says it

The chat panel is integrated with the GraphRAG pipeline (section 3) with conditional retrieval. Following the cognitive forcing function (CFF) design, the chatbot first shows an Answer Context Summary panel (Figure 3-A) before giving the final answer.

The graph panel displays a node-link diagram where entities are displayed in columns following the link specified in FoodAtlas. p. 4
Where the paper says it

2 Graph Panel The graph panel contains a node-link diagram displaying Disease, Chemical, and Food entities, as well as the relationships between them. Entities are displayed in columns, with the disease column on the left, chemical column in the center, and food column on the right.

This visualization supports progressive disclosure by starting with a high-level textual summary and allowing users to progressively explore the hierarchy on-demand. p. 5
Where the paper says it

VArify is designed to display information in various levels of detail following the principle of progressive disclosure. First, the user begins with a high-level textual summary of the retrieved context via the context summary panel.

The case study demonstrates how Alice uses VArify to verify and iterate upon GraphRAG-enhanced LLM answers. Alice successfully identifies a knowledge graph construction error regarding pineapple containing ozone and positively correlated with pulmonary heart disease by inspecting the metadata card and source literature. p. 6

Where the paper says it

Reading the text shows that pineapple is simply “treated with ozone”, suggesting a knowledge graph construction error. Since no other references are shown for this pineapple-ozone relationship, she decides that there is no evidence that pineapple contains ozone and, in turn, is positively correlated with heart disease.

Furthermore, Alice identifies a retrieval error where heart disease was mapped to pulmonary heart disease and corrects it by selecting an alternative entity mapping. This iterative cross-referencing between LLM responses and supporting evidence empowers users to calibrate trust based on grounded facts. p. 7
Where the paper says it

For example, P2 found that honey is not in the graph view, and concluded that “this question is generated by, not based on the graph, but based on the prior knowledge”. Ultimately, this iterative cross-referencing between the LLM responses and supporting evidence empowers users to calibrate trust based on grounded facts, fostering a more critical and informed trust in the system.

The authors discuss design opportunities for CFFs as informative functions that offer cognitive rewards rather than just interruptions. They also propose using visualization as a nudge, where the view separation between the chat and graph panels steers users toward verification. p. 7

Where the paper says it

In our system, the layout of the chatbot and the graph panel is such a nudge that steers users toward verification. The view separation reminds the user of the distinction between context and LLM responses, and the hyperlinks connecting the chat panel and the visualization provide easy access to the underlying context.

Finally, they address the challenge of unknown unknowns, noting that verification is limited when relevant knowledge is not retrieved by GraphRAG or known to the expert. They also suggest dynamic grouping of entities to support personalized sensemaking based on shifting task-based hierarchies. p. 8
Where the paper says it

A static ontology that only offers a “Chemical Family” view fails to support these shifting task-based hierarchies, making it difficult for the user to verify if the LLM has accounted for all relevant functional components. This suggests that a dynamic grouping of entities that can be adjusted on the fly through natural language is necessary to support personalized sensemaking and verification of retrieved context.

The user study results showed that participants actively engaged in verification behaviors and reported a notable perceived need for verification. Participants preferred the clear visual distinction between LLM responses and retrieved context as it served as a constant reminder of potential discrepancies. p. 6

Where the paper says it

The clear visual distinction between LLM responses and retrieved context was preferred, as it served as a constant reminder that discrepancies might exist between the two views. Participants reported a high sense of responsibility in judging answer correctness (M = 5.

The study concluded that by providing a clear visualization for retrieved context, the system effectively empowers domain experts to maintain agency and critical oversight when interacting with GraphRAG chatbots. p. 6
Where the paper says it

11, SD = 1. This suggests that by providing a clear visualization for retrieved context, the system effectively empowers domain experts to maintain agency and critical oversight when interacting with GraphRAG chatbots.

The paper concludes that verification supported by visualizing retrieved context is a process of aligning the domain knowledge of expert users with the context retrieved from the knowledge graph. The ultimate goal of verification should include knowing the limits of verifiability, as this prompts users to remain vigilant even when no explicit signals of failure are provided. p. 8

Where the paper says it

, medical and legal AI), the ultimate goal of verification should not only include being “right”, but also knowing the limits of verifiability. Consequently, designing for revealing limits of verifiability becomes as important as designing for accuracy, as it prompts users to remain vigilant even when the system provides no explicit signals of failure.

The work suggests that visualization appears to be a promising direction for human-centered verification approaches that empower users with greater agency. p. 8
Where the paper says it

Nevertheless, visualization appears to be a promising direction to pursue for its inherent human-centered focus and immense design opportunities. We hope this direction is continued with more focus on human-centered verification approaches that empower users with greater agency.

The study found that participants treated the chat and graph as complementary resources, with the graph view providing more specific but potentially incomplete information compared to general LLM responses. Participants demonstrated strategies like breaking down questions into sub-questions to independently verify information in each step, reducing the amount of relevant nodes that need inspection. p. 7

Where the paper says it

P5 approached the quizzes in a way that helps reduce the complexity of verification: instead of directly looking for an answer from the chatbot, P5 first broke down the question into multiple sub-questions, asked each question in sequence, and combined all collected information without relying on the chatbot. This allows P5 to independently verify information in each sub-question, reducing the amount of relevant nodes and relationships that need inspection.

The findings underscore the limitations of GraphRAG, as technical errors during automated construction process can surface issues like an improbable relationship between blueberries and curcumin. p. 7
Where the paper says it

For example, “fried chicken” is not a food entity in the knowledge graph, despite its clear importance when considering food-disease relationships. Technical errors during the automated construction process also surfaced, such as an improbable “contains” relationship between blueberries and curcumin.

The work suggests that users can leverage visualization to identify representative error cases and reverse-engineer them back into generalizable patterns for systematic corrections. p. 8
Where the paper says it

Users can first leverage the visualization to identify representative error cases, and then reverse-engineer them back into generalizable patterns that can be incorporated into the construction pipeline. This approach facilitates systematic corrections with a human-in-the-loop framework.

The study utilized a relatively small sample size, and the current backend architecture does not incorporate the most recent advancements in GraphRAG state-of-the-art. The findings may not account for varying hallucination rates or reasoning behaviors inherent in other LLMs, as the evaluation was restricted to a single model family (GPT-4o). p. 8

Where the paper says it

The current backend architecture does not incorporate the most recent advancements in GraphRAG state-of-theart, and our evaluation was restricted to a single model family (GPT-4o). Consequently, the findings may not account for the varying hallucination rates or reasoning behaviors inherent in other LLMs.

The study did not explore the integration of autonomous agents equipped with iterative reasoning capabilities and real-time web-browsing support. p. 8
Where the paper says it

Consequently, the findings may not account for the varying hallucination rates or reasoning behaviors inherent in other LLMs. Furthermore, we did not explore the integration of autonomous agents equipped with iterative reasoning capabilities and real-time web-browsing support, which have become standard in public-facing chatbots but were outside the scope of this controlled study.

The work also did not explore ethical dimensions of human-AI collaboration, such as the shifting nature of accountability in safety-critical domains. p. 8
Where the paper says it

While our tasks were designed to simulate expert inquiry, they represent a simplified version of the multi-stage practices typical in the field. Additionally, this work did not explore ethical dimensions of human-AI collaboration, such as the shifting nature of accountability that might be of high priority in safety-critical domains.

The paper offers a promising direction for visualization due to its inherent human-centered focus and immense design opportunities. The system's ability to facilitate verification by enabling the identification of problematic entities while providing clear insights into the GraphRAG retrieval process is demonstrated through the case study. The overall contribution lies in showing how visualization can be used to support verification without causing over-trust. p. 1

Where the paper says it

Our work targets domain experts with limited prior experience with knowledge graphs. As a result, instead of proposing novel visualization techniques for query construction or query results, we aim at designing visualizations that support verification without causing over-trust.

The paper aims to continue this direction with more focus on human-centered verification approaches that empower users with greater agency. p. 8
Where the paper says it

Nevertheless, visualization appears to be a promising direction to pursue for its inherent human-centered focus and immense design opportunities. We hope this direction is continued with more focus on human-centered verification approaches that empower users with greater agency.

The authors suggest that CFFs may be most sustainable when they offer immediate cognitive “rewards,” such as evidence synthesis, rather than acting solely as interruptions. They propose that visualization functions as a nudge by using view separation and hyperlinks to steer users toward critical evaluation of AI outputs. p. 7

Where the paper says it

2 Visualization as a nudge Nudge Theory has been proposed as a non-intrusive way to steer cognitive processes and mitigate overreliance on AI [48]. Unlike “hard” constraints that force specific workflows such as CFFs, nudges introduce soft interventions that are typically implemented through subtly designed “choice architecture” [43], in which every interaction choice made by the user is thoughtfully architected to steer users toward a more critical evaluation of AI outputs.

The paper also suggests that dynamic grouping of entities can be necessary to support personalized sensemaking by adapting hierarchies on the fly through natural language. p. 8
Where the paper says it

A static ontology that only offers a “Chemical Family” view fails to support these shifting task-based hierarchies, making it difficult for the user to verify if the LLM has accounted for all relevant functional components. This suggests that a dynamic grouping of entities that can be adjusted on the fly through natural language is necessary to support personalized sensemaking and verification of retrieved context.

The work suggests that rule-based refinement or few-shot prompting strategies can address limitations in knowledge graph construction through a human-in-the-loop framework. p. 8
Where the paper says it

Our work can be extended with rule-based refinement [34] or few-shot prompting strategies [31] to address this issue. Users can first leverage the visualization to identify representative error cases, and then reverse-engineer them back into generalizable patterns that can be incorporated into the construction pipeline.

The system supports explicit linkage between LLM responses and the graph visualization by converting FoodAtlas IDs to entity hyperlinks, which speeds up alignment when looking for a single entity. The hierarchical visualization is inspired by file directory browsers, using breadcrumbs and indentations to support navigation along the path. p. 4

Where the paper says it

The user can also use the “Locate” buttons to navigate to each related entity and back to the selected entity. Hierarchy construction and controls The hierarchical visualization is inspired by file directory browsers.

The system supports context editing by allowing users to delete entity groups or edit subqueries directly via a context control panel. p. 5
Where the paper says it

A context control panel is provided at the top left of the graph panel that contains the necessary information about the subqueries. In this panel, users can highlight entities and links associated with a subquery, edit the entity mapping under each subquery, or delete a subquery, which removes all associated entities and relationships altogether.

This allows for progressive disclosure of information, starting with a high-level summary and moving to detailed inspection of specific entities and their metadata cards. p. 5
Where the paper says it

This system of progressive disclosure lends itself to progressive verification. The user first begins by verifying the GraphRAG context via the context summary panel and the graph panel; if needed, they can then verify the underlying knowledge graph itself using relevant entities’ metadata cards.

The system successfully regenerates the LLM answer after context edits using a “Requery With Changes” button. p. 6
Where the paper says it

Finally, Alice clicks on the “Requery With Changes” button above the Minimap (Figure 3-B) to apply her context updates and regenerate a new LLM answer. The system returns a new context Summary which mentions various foods like fruits, vegetables, spices, and nuts, and states that the retrieved context “encompasses a wide range of food products”.

The final result is a new context Summary which can be used to reveal the full LLM response, allowing users to calibrate trust based on grounded facts. p. 7
Where the paper says it

For example, P2 found that honey is not in the graph view, and concluded that “this question is generated by, not based on the graph, but based on the prior knowledge”. Ultimately, this iterative cross-referencing between the LLM responses and supporting evidence empowers users to calibrate trust based on grounded facts, fostering a more critical and informed trust in the system.

p. 7
Where the paper says it

For example, P2 found that honey is not in the graph view, and concluded that “this question is generated by, not based on the graph, but based on the prior knowledge”. Ultimately, this iterative cross-referencing between the LLM responses and supporting evidence empowers users to calibrate trust based on grounded facts, fostering a more critical and informed trust in the system.

The paper's primary contribution is demonstrating how visualization can be used to support verification without causing over-trust. The system's design supports progressive disclosure of graph and domain-specific information to prevent cognitive overload while supporting progressive verification. The study confirms that by providing a clear visualization for retrieved context, the system effectively empowers domain experts to maintain agency and critical oversight. p. 6

Where the paper says it

11, SD = 1. This suggests that by providing a clear visualization for retrieved context, the system effectively empowers domain experts to maintain agency and critical oversight when interacting with GraphRAG chatbots.

The paper provides a framework for designing CFFs that offer immediate cognitive “rewards,” transforming them into informative functions rather than mere interruptions. p. 7
Where the paper says it

Throughout the course of answering three quizzes, participants consistently engaged with the panel’s content rather than bypassing it. While these results are not definitive proof that habituation is impossible in longer-term deployments, they suggest a promising design opportunity: CFFs may be most sustainable when they offer immediate cognitive “rewards”, such as evidence synthesis, rather than acting solely as interruptions.

The findings highlight the importance of addressing the limitations of knowledge graphs through human-in-the-loop refinement strategies. p. 8
Where the paper says it

5 Addressing limitations of knowledge graphs The user study highlighted several limitations of the underlying knowledge graph, including errors arising from the automated construction process. Given that knowledge graph construction involves integrating vast, heterogeneous data sources [28], human validation remains a practical necessity.

The paper ultimately advocates for continued focus on human-centered verification approaches that empower users with greater agency. p. 8
Where the paper says it

Nevertheless, visualization appears to be a promising direction to pursue for its inherent human-centered focus and immense design opportunities. We hope this direction is continued with more focus on human-centered verification approaches that empower users with greater agency.

The authors successfully demonstrated how VArify facilitates verification by enabling the identification of problematic entities while providing clear insights into the GraphRAG retrieval process and subquery sources. The system's ability to facilitate verification is further shown through Alice's iterative process where she removes irrelevant subqueries and corrects entity mappings to ensure context alignment. This process leads to a new context Summary that encompasses a wider range of relevant foods and chemicals, leading Alice to trust the final LLM response more. p. 6

Where the paper says it

Finally, Alice clicks on the “Requery With Changes” button above the Minimap (Figure 3-B) to apply her context updates and regenerate a new LLM answer. The system returns a new context Summary which mentions various foods like fruits, vegetables, spices, and nuts, and states that the retrieved context “encompasses a wide range of food products”.

The overall gain is the creation of an iterative information-seeking experience where users transition from passive consumers of LLM responses to active investigators leveraging different system components to triangulate their expertise with provided data. p. 7
Where the paper says it

This allows P3 to collect more relevant information about gout. By employing these strategies, participants effectively transitioned from passive consumers of LLM responses to active investigators, leveraging different system components to triangulate their expertise with the provided data.

p. 7
Where the paper says it

This allows P3 to collect more relevant information about gout. By employing these strategies, participants effectively transitioned from passive consumers of LLM responses to active investigators, leveraging different system components to triangulate their expertise with the provided data.

The paper provides a framework for designing visualizations that support domain-aligned reasoning by visualizing multi-hop relationships through bundled edges between groups. p. 3
Where the paper says it

DR2 : Support structured and domain-aligned reasoning. F2 shows that users rely on surface-level matching within tabular representations, while F3 shows that users struggle to reconstruct multi-hop relationships, often leading to shallow or incorrect verification.

p. 3
Where the paper says it

DR2 : Support structured and domain-aligned reasoning. F2 shows that users rely on surface-level matching within tabular representations, while F3 shows that users struggle to reconstruct multi-hop relationships, often leading to shallow or incorrect verification.

The study found that participants often rely on surface-level matching within tabular representations rather than reconstructing multi-hop relationships, which is a limitation of the current visualization design. The paper highlights that static ontologies fail to support shifting task-based hierarchies, necessitating dynamic grouping that can be adjusted on the fly through natural language for personalized sensemaking. p. 8

Where the paper says it

A static ontology that only offers a “Chemical Family” view fails to support these shifting task-based hierarchies, making it difficult for the user to verify if the LLM has accounted for all relevant functional components. This suggests that a dynamic grouping of entities that can be adjusted on the fly through natural language is necessary to support personalized sensemaking and verification of retrieved context.

The authors suggest that uncertainty visualization could be leveraged by alerting users when context relevancy is low to encourage verification. p. 7
Where the paper says it

In the case of GraphRAG, it might be feasible to estimate uncertainty through context relevance. While we did not explore such designs in our system, prior research has shown that even simple designs, such as alerting users when context relevancy is low, or highlighting responses derived from uncertain context, could encourage verification [37].

The ultimate goal is to design for revealing limits of verifiability, acknowledging that relevant knowledge not retrieved by GraphRAG might simply be unknown to the expert. p. 8
Where the paper says it

, medical and legal AI), the ultimate goal of verification should not only include being “right”, but also knowing the limits of verifiability. Consequently, designing for revealing limits of verifiability becomes as important as designing for accuracy, as it prompts users to remain vigilant even when the system provides no explicit signals of failure.

The paper suggests that users can leverage visualization to identify representative error cases and reverse-engineer them back into generalizable patterns for systematic corrections. p. 8
Where the paper says it

Users can first leverage the visualization to identify representative error cases, and then reverse-engineer them back into generalizable patterns that can be incorporated into the construction pipeline. This approach facilitates systematic corrections with a human-in-the-loop framework.

The authors conclude that verification supported by visualizing retrieved context is a process of aligning the domain knowledge of expert users with the context retrieved from the knowledge graph. p. 8
Where the paper says it

3 Handling unknown unknown remains a challenge The verification supported by visualizing retrieved context is a process of aligning the domain knowledge of expert users with the context retrieved from the knowledge graph. Figure 5 shows the four possible scenarios of alignment.

The system's ability to facilitate verification is demonstrated through Alice's iterative process where she removes irrelevant subqueries and corrects entity mappings to ensure context alignment. The final result is a new context Summary which can be used to reveal the full LLM response, allowing users to calibrate trust based on grounded facts. This iterative process leads to a new context Summary that encompasses a wider range of relevant foods and chemicals, leading Alice to trust the final LLM response more. p. 6

Where the paper says it

Finally, Alice clicks on the “Requery With Changes” button above the Minimap (Figure 3-B) to apply her context updates and regenerate a new LLM answer. The system returns a new context Summary which mentions various foods like fruits, vegetables, spices, and nuts, and states that the retrieved context “encompasses a wide range of food products”.

The paper provides a framework for designing visualizations that support domain-aligned reasoning by visualizing multi-hop relationships through bundled edges between groups. The overall gain is the creation of an iterative information-seeking experience where users transition from passive consumers of LLM responses to active investigators leveraging different system components to triangulate their expertise with provided data. The paper provides a framework for designing visualizations that support verification without causing over-trust. p. 1
Where the paper says it

Our work targets domain experts with limited prior experience with knowledge graphs. As a result, instead of proposing novel visualization techniques for query construction or query results, we aim at designing visualizations that support verification without causing over-trust.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Cognitive Forcing Functions for Verification: Implement cognitive forcing functions such as delaying answer reveal, requiring user interaction before showing the response, or prompting users to verify specific claims to interrupt habitual reliance on the first response and encourage deliberate verification, transforming the panel into an informative forcing function rather than just an interruption.

  2. Context Summary Panel for Bias Mitigation: Utilize a context summary panel that describes retrieved data in a few bullet points before revealing the final answer, serving as a critical cognitive buffer that steer[s] users away from anchoring their verification judgement on the LLM’s responses.

  3. Entity Hyperlinking for Alignment: Convert FoodAtlas IDs in LLM responses to entity hyperlinks that automatically center the graph panel on the relevant entity, which greatly speeds up the alignment process, especially when looking for a single entity in a large retrieved graph context.

  4. Progressive Disclosure Visualization: Design the graph panel to display information in various levels of detail following the principle of progressive disclosure, starting with a high-level textual summary and then moving to aggregate views before allowing users to progressively explore the hierarchy on-demand.

  5. Dynamic Grouping for Personalized Sensemaking: Implement dynamic grouping of entities that can be adjusted on the fly through natural language to support personalized sensemaking, allowing experts to discern whether a discrepancy stems from a retrieval error or a fundamental mismatch of conceptual models.

  6. Uncertainty Visualization: Explore methods to quantify uncertainty in LLM responses by alerting users when context relevancy is low or highlighting responses derived from uncertain context, which could encourage verification even when the system provides no explicit signals of failure.

  7. Rule-Based Knowledge Graph Refinement: Extend the knowledge graph construction pipeline with rule-based refinement or few-shot prompting strategies to address limitations in automated construction, allowing users to reverse-engineer them [error cases] back into generalizable patterns that can be incorporated into the construction pipeline.

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) grounds Large Language Model (LLM) responses in structured, domain-specific knowledge. However, verification remains difficult because errors can arise at multiple layers: an answer may not be faithful to its retrieved context, the retrieval pipeline may select irrelevant evidence, and the knowledge graph may omit relevant facts or contain incorrect triplets. We present VArify, a visual analytics system for food-science researchers to inspect and repair these layers in a GraphRAG workflow. A formative study with five participants using a conventional chat-and-table interface revealed behaviors reflecting answer anchoring, over-trust, and difficulty tracing multi-hop relationships. In response, VArify combines an evidence-first answer reveal with coordinated chat and hierarchical graph views, provenance inspection, and subquery editing. It connects the generated answer, the retrieval state, and the provenance-linked graph so that researchers can trace discrepancies across them and repair the retrieved context. From a qualitative system-use study on six food science researchers, we characterize how they scrutinize GraphRAG evidence with progressive visual inspection. Our study also reveals an important verification boundary: information absent from both retrieval and a user's knowledge remains unverifiable. We discuss the broader role of visualization in supporting oversight of LLM responses and outline future research directions.

Sources

Related papers