Source-linked AI summary
Graphionale: How Graph Visualizations of LLM Rationales Affect Human Decision Making
Xinru Wang, Zhexuan Ma, Ming Yin, Shuai Ma, Thomas W Malone
TL;DR
Dense textual LLM rationales can burden users and obscure the relational structure needed to inspect reasoning. This paper develops Graphionale, an interactive argument-map-style rationale representation, and evaluates it against textual rationales across task modalities and difficulty levels. Graphical rationales improve trust calibration for verbal reasoning but impair it for visual reasoning, while users prefer the opposite formats in each modality.
Problem
Dense, linear LLM rationales obscure multi-hop reasoning, while evidence remains limited on how graphical rationales affect human understanding and decision making.
Method
Graphionale combines argument mapping and text condensation, and a controlled online study compares graphical and textual rationales across verbal and visual reasoning and question difficulty.
Results
Graphical rationales improve trust calibration for verbal reasoning but impair it for visual reasoning, while subjective preferences reverse across modalities.
Takeaways & Limitations
Rationale format should be matched to task modality because the format supporting better-calibrated decisions is not the one users prefer.
Takeaways & Limitations
Findings are based on well-structured QA decision tasks and one argument-map-style representation with condensed node text, limiting generalizability.
Abstract
from arXiv · showhide
Large Language Models (LLMs) are increasingly equipped with augmented reasoning capabilities to generate rationales that support human decision-making. Yet these text-dense rationales often impose substantial cognitive burdens. Building on a formative co-design study that identified user preferences for non-linear reasoning representations, we developed Graphionale as a testbed for empirically studying argument-map-style rationale visualization. This system transforms linear LLM rationales into interactive, multi-level graphs. It explicitly structures logical relationships (e.g., conclusions, premises, support, and objections), while further extracting entities and relations within each statement to construct condensed node-link representations. We conduct a large-scale online user study (N = 204) to examine when graphical rationales are more effective than textual ones, across varying task modality (verbal vs. visual reasoning), rationale format (textual vs. graphical), and question difficulty (easy vs. hard). Our results show that graphical rationales do not help uniformly: they improve trust calibration for verbal reasoning yet feel more cognitively demanding and less satisfying; for visual reasoning, they impair calibration yet feel more engaging and helpful. In each modality, the format that better supports calibrated decisions is not the one users prefer, highlighting that matching rationale format to task modality is key to effective AI explanation design. Our findings contribute empirical design knowledge about when and how graphical rationales support human decision making, and inform the next-generation reasoning-aware AI interfaces.
1 Introduction
The paper addresses the difficulty of evaluating dense, linear LLM rationales by developing Graphionale, an argument-map-style graphical rationale system and testing when it supports decision making. A controlled study finds that graphical rationales have modality-dependent effects, improving calibration for verbal reasoning but impairing it for visual reasoning while producing opposite preference patterns.
- Motivation: Dense textual rationales make users reconstruct branching, aggregating, and eliminative reasoning from sequential sentences.The paper motivates graphical representations as a way to expose relational structures that linear text does not intuitively represent.
- Research gap: Design principles for graphical LLM rationales remain underexplored, with limited evidence about their effects on understanding and decision making.The paper therefore investigates both how such representations should be designed and how they compare with textual rationales.
- Approach: Graphionale combines argument mapping with text condensation to create glanceable graphical representations of LLM rationales.The system structures reasoning and compresses node-level text while serving as the prototype for empirical comparison.
- Study: N = 204 participants compared graphical and textual rationales across verbal versus visual reasoning and easy versus hard questions.The controlled online study examined both decision-making outcomes and user perceptions across these contexts.
- Findings: Graphical rationales improved trust calibration for verbal reasoning but impaired it for visual reasoning, while user preferences reversed across modalities.For verbal reasoning, graphs were more cognitively demanding and less satisfying; for visual reasoning, they were more engaging and helpful.
2 Related Work
Prior work uses graphs for AI explanation, structured reasoning, and interactive LLM exploration, but graphical LLM rationales remain insufficiently designed and evaluated. This paper positions its contribution at the human–AI interface, where dense text can obscure multi-hop reasoning and constrain efficient evaluation.
- Graph-based AI explanation: Knowledge graphs and related structures have supported human-readable reasoning chains, causal reasoning, and counterfactual reasoning in AI systems.These systems represent relationships among entities or decisions to expose reasoning structure.
- Structured LLM reasoning: LLM reasoning research has expanded from linear Chain-of-Thought toward Tree-of-Thought and Graph-of-Thought approaches that explore multiple branches.These approaches concern model reasoning augmentation rather than necessarily presenting rationales to human users.
- Cognitive grounding: Graph-like representations align with human associative knowledge structures and can help manage cognitive load through perceptual organization and selective detail.Concept maps and related representations exploit visual organization while recognizing limited visual working memory.
- Argument mapping: Argument maps explicitly link premises, evidence, and conclusions to clarify reasoning, identify weaknesses, and support critical evaluation.Earlier navigable-graph systems also used graph structures to help synthesize complex information and support deliberation.
- LLM interfaces: Existing LLM graph interfaces primarily support output exploration or organization, while rationale-focused systems have only recently begun modeling implicit reasoning dependencies.Prior systems include interactive diagrams, multilevel organization, provenance graphs, and reasoning-trace visualizations.
- Research gap: Limited empirical evidence remains on how graphical LLM rationales should be designed and how they affect understanding, decisions, and human–AI collaboration.This gap motivates systematic evaluation of graphical rationale design and user outcomes.
3 Formative Study: A Co-Design Exploration
A formative co-design study used participant interaction, comparison, sketching, and thematic analysis to derive design goals for graphical LLM rationales. Participants emphasized glanceability, alignment with decision logic, clear visual semantics, and bounded complexity.
- 3.1 Participants and Study Procedure: Ten LLM-familiar participants completed co-design activities to express preferences for graphical rationale representations.The sample included graduate students, research scientists, and a software engineer.
- 3.1 Participants and Study Procedure: A low-fidelity interface let participants compare textual and graphical rationales derived from CommonsenseQA explanations.The graphical representations used nodes for premises and conclusions and directed edges for inference relationships.
- 3.1 Participants and Study Procedure: Participants reviewed two questions, described difficulties, compared formats, proposed improvements, and sketched alternatives during approximately 25-minute sessions.Sessions were recorded for subsequent analysis.
- 3.2 Analysis and Results: Thematic analysis identified four themes describing how graphical LLM rationales should be improved.Two authors iteratively refined a shared codebook during analysis.
- Glanceable nodes: Participants wanted compressed keyword-based nodes with structured text available as a fallback when diagrams failed to improve clarity.This design goal targets fast scanning without eliminating access to fuller explanations.
- Decision logic: Participants preferred maps that mirror human decision logic, especially criteria, comparison, elimination, and summary across answer options.They also proposed adapting map depth to task difficulty and avoiding irrelevant terms.
- Visual semantics: Clear visual semantics should make support, rejection, grouping, and reasoning strength interpretable at a glance.Participants suggested colors, icons, or symbols and clearer edge semantics.
- Bounded complexity: Complexity should be bounded through node merging, depth limits, and progressive disclosure, with a suggested scale of about three to four layers and three to five nodes per layer.Participants reported that large diagrams created navigation costs and feelings of being overwhelmed.
4 Graphionale Interface
Graphionale transforms LLM rationales into interactive argument graphs that expose reasoning relationships while condensing node text. Its design combines structured argument mapping, progressive disclosure, and interactive navigation to support glanceable inspection.
- Interface concept: Graphionale presents a logic-puzzle rationale as a graph that organizes statements and their relationships instead of requiring paragraph-by-paragraph reconstruction.The scenario illustrates the contrast between traditional textual presentation and graphical organization.
- Argument structure: The graph aggregates constraints, eliminates candidate time ranges, and leads to a single valid interval.This structure mirrors a decision process based on combining evidence and removing alternatives.
- Progressive disclosure: Graph complexity is progressively disclosed by showing high-level constraints first and expanding details on demand.This keeps the overview bounded while allowing access to deeper reasoning.
- Node representation: Nodes use compressed keyword-based statements, including subject–predicate–object triplets, for quick scanning.For example, a time range and an activity are represented as a structured relation within a node.
- Relational encoding: Edges encode relationships such as support and objection between statements.Argument-map nodes represent premises, sub-conclusions, and final conclusions, while edge styling distinguishes relation types.
- Generation pipeline: Graphical rationale generation separates argument-map construction from compression of statement text within each node.The two-step decomposition defines the core rationale-generation pipeline.
- Implementation: The full-stack application renders questions and rationales, manages generation and derivation of both views, and persists task data in PostgreSQL.Its graphical view supports zooming, viewport navigation, branch expansion and collapse, and on-demand full-text display.
5 Technical Evaluation
The technical evaluation assessed whether generated graphical rationales were structurally valid, semantically coherent, and useful for detecting reasoning errors. Overall graph structure was highly reliable, but correct and incorrect rationales exhibited distinct reasoning-quality problems.
- Generation quality: Graph generation quality was high, with only 5 structural issues among 180 rationales (2.8%).All issues involved excessive edges in BBH Logical Deduction questions; no violations occurred in node scaling, node count, text length, or consistency.
- Correct-answer rationales: Correct-answer rationales contained redundant reasoning steps in 15/120 cases (13%) and unverifiable visual claims in 10/120 cases (8%).Redundant intermediate nodes particularly affected I-RAVEN tasks, while unverifiable claims concerned subtle visual attributes.
- Incorrect-answer rationales: Incorrect-answer rationales showed missing evidence in 10/60 cases (17%), problematic evidence in 26/60 cases (43%), and incorrect conclusions in 30/60 cases (50%).Problematic evidence was especially prominent in I-RAVEN, whereas incorrect inferences were most prominent in BBH.
- Diagnostic value: The three error patterns were almost entirely absent from correct-answer rationales, supporting their diagnostic value for identifying LLM reasoning errors.Two coders evaluated graph structure, semantic reasoning, and functional correctness, and inter-rater reliability was high (Cohen’s κ= 0.94).
6 User Study Design
The user study compared graphical and textual rationales across verbal and visual reasoning tasks, with easy and hard questions within each condition. A 2 × 2 between-subjects design and trial-level trust-calibration measures enabled evaluation across decision-making contexts.
- Design: The 2 × 2 between-subjects design crossed rationale format with task modality, while each condition included both easy and hard questions.The four conditions were BBH × Graph, BBH × Text, I-RAVEN × Graph, and I-RAVEN × Text.
- Participants and procedure: 204 valid participants completed the study after attention checks, exceeding the power-analysis requirement of 171 participants.Participants were randomly assigned to conditions and completed a pre-study survey, tutorial, 15-trial task phase, and post-study survey.
- Tasks: The study used BBH for verbal reasoning and I-RAVEN for visual reasoning, selecting one easy and two hard task categories in each dataset.Thirty questions were sampled per category, and incorrect-answer trials were included to test whether rationales helped participants detect reasoning errors.
- Design: The study examined graphical versus textual rationales across verbal and visual reasoning, easy and hard questions, and participants’ need for cognition.Rationale format and task modality were between-subjects factors; question difficulty was manipulated within subjects, while NFC was analyzed as a moderator.
- Measures: Trust calibration was measured through Appropriate Trust, Over-Trust, Under-Trust, and Error Correction.Subjective outcomes included usability, task load, satisfaction, engagement, helpfulness, understanding, trust, critical thinking, and learning gain.
- Analysis: Objective outcomes were analyzed with participant-level random-intercept linear mixed-effects models, while subjective outcomes used 2 × 2 ANOVA and Welch’s t-tests.NFC moderation and thematic analysis of open-ended responses were conducted as secondary analyses.
7 Results
Graphical rationale effects reversed across task modalities: they improved objective trust calibration for verbal reasoning but impaired it for visual reasoning. Subjective preferences showed the opposite pattern, with graphs feeling worse in verbal tasks and better in visual tasks.
- Objective outcomes: The rationale format × task modality interaction was significant across all four trust-calibration metrics, with format effects reversing between verbal and visual reasoning.The interaction was significant for Appropriate Trust (b= 0.049), Over-Trust (b= −0.064), Under-Trust (b= −0.040), and Error Correction (b= 0.090).
- Objective outcomes: In verbal reasoning, graphical rationales reduced Over-Trust by Δ = −0.144 and increased Error Correction by Δ = +0.127 after collapsing across difficulty.On hard questions, Over-Trust decreased by Δ = −0.199 and Error Correction increased by Δ = +0.182; these benefits were absent on easy questions.
- Objective outcomes: In visual reasoning, graphical rationales reduced Appropriate Trust by Δ = −0.165, increased Under-Trust by Δ = +0.186, and reduced Error Correction by Δ = −0.233.The calibration disadvantage was present at both difficulty levels, indicating that difficulty did not explain the visual-reasoning effect.
- Subjective experience: In verbal reasoning, graphical rationales lowered Satisfaction (Δ = −0.76), Usability (Δ = −0.72), and Learning Gain (Δ = −0.82), while increasing Task Load by Δ = +0.53.Thus, the format that improved calibration was experienced as more demanding and less satisfying.
- Subjective experience: In visual reasoning, graphical rationales increased Satisfaction by Δ = +0.74, Engagement by Δ = +0.99, and Helpfulness by Δ = +0.68, while increasing Task Load by Δ = +0.94.Participants therefore preferred graphs in the modality where objective calibration was impaired.
- Individual differences: Need for Cognition moderated objective calibration effects, with the verbal-reasoning benefit concentrating among High-NFC participants while visual impairment appeared across NFC groups.Subjective outcomes showed no reliable continuous NFC moderation, making subgroup rating differences exploratory.
- Design features and qualitative themes: Participants rated all four graphical features above the scale midpoint, led by Progressive Disclosure (M= 5.66) and Semantic Color-Coding (M= 5.58).Qualitative responses valued stepwise structure but described different uses across modalities: logic tracing for BBH and visual-pattern decomposition for I-RAVEN.
- Summary: The paper attributes the modality reversal to graphs exposing errors in verbal reasoning but lacking sufficient perceptual support for reliable verification in visual reasoning.In visual tasks, participants requested attribute tracks, legends, and direct highlighting in source images.
8 Discussion
Graphical rationales supported calibrated decisions only when their representation matched the task, while subjective preferences often favored the opposite format. The discussion highlights cognitive demands, user engagement, graph complexity, and boundaries on generalizability.
- Objective–Subjective Dissociation: Graphical rationales improved trust calibration for verbal reasoning but impaired it for visual reasoning, while subjective ratings reversed across modalities.They were more demanding and less satisfying for verbal reasoning, but more engaging, helpful, and satisfying for visual reasoning.
- Modality Mismatch: Verbal reasoning benefited when argument-graph nodes mapped to logical inferences, whereas visual reasoning required perceptual grounding absent from the abstract graph.I-RAVEN participants had to maintain the source matrix while parsing propositional nodes, creating extraneous cognitive load.
- Objective–Subjective Dissociation: 28.2s vs. 36.7s: visual-reasoning participants spent less time with graphical than textual rationales, despite rating graphical rationales as more demanding.Verbal reasoning showed no format effect on response time, reinforcing the dissociation between objective behavior and subjective judgments.
- User Engagement and Difficulty: Graphical rationales improved verbal calibration primarily for high-NFC participants, suggesting benefits required effortful engagement with the argument scaffold.The pattern is consistent with central-route processing, which depends on ability and motivation.
- User Engagement and Difficulty: The graph advantage emerged exclusively on hard questions, where explicit structure provided scaffolding for deeper inference.This pattern aligns with prior findings that explanation benefits increase when tasks are sufficiently challenging.
- Design Implications: Graph complexity may follow an inverted-U relationship with calibration benefit, motivating pruning of decision-critical nodes and progressive disclosure.Progressive disclosure can manage branching complexity by collapsing branches on demand.
- Design Implications: Visual-task explanations should anchor directly in source visuals while preserving argument structure, and future systems could adapt formats through hybrid representations.The proposed split pairs graph-structured arguments with verbal reasoning and spatial annotations with visual reasoning.
- Limitations and Future Work: The findings are constrained by well-structured decision-making tasks, limited coverage of open-ended collaboration, and one argument-map-style implementation.Visual-reasoning results may not generalize to other visual tasks or graphical rationale forms.
9 Conclusion
The conclusion presents Graphionale as a prototype for interactive argument-graph rationales and summarizes its controlled evaluation. Its central finding is that calibrated decision support and user preference diverged across verbal and visual reasoning.
- Conclusion: Graphionale transforms linear LLM rationales into interactive argument graphs by combining argument mapping, text condensation, and progressive disclosure.The system was grounded in a formative co-design study.
- Conclusion: N = 204: a controlled user study found that graphical rationales improved verbal calibration but impaired visual calibration, while user preferences favored the opposite formats.Verbal graphs reduced over-trust and increased error correction, especially on hard questions and among analytically engaged users.
- Conclusion: Argument maps structure rationales by classifying claims as premises, sub-conclusions, or conclusions and connecting them with directed support relationships.The prototype’s formative interface compared textual and graphical rationales after participants reviewed the textual version.
- Conclusion: Participant sketches grouped correct and incorrect options into branches, used shared explanations for distractors, and represented elimination through explicit non-support edges.The sketches also used hierarchical nodes and progressive unfolding.
B System Implementation Details
Graphionale’s implementation converts question–answer text into structured reasoning graphs, compresses node labels, and serializes the result in a typed schema. Dataset-specific prompts constrain reasoning for I-RAVEN and Big-Bench Hard.
- Reasoning Structure: The first pipeline stage transforms a question–answer pair into an argument map, a structured reasoning graph.The implementation uses a dedicated prompt for this conversion.
- Reasoning Structure: Each node expresses one atomic claim as a complete sentence and is classified as a premise, sub-conclusion, or conclusion.Chained reasoning steps and independent entities are split into separate nodes.
- Edge Construction: Edges encode either support or objection, including objection chains for eliminated candidates and support links for surviving candidates.Multiple premises receive separate edges to the same target, and sub-conclusions can form deeper chains.
- Output Schema: The output schema stores node identifiers, labels, optional SVGs, node types, and edges with source, target, and relation fields.Relations are serialized as support or objection.
- Node Compression: Compressed nodes use subject–predicate–object triplets when natural, otherwise retaining a fallback label of at most six words.The compression preserves concrete attributes, values, choice numbers, and rule names.
- Dataset-Specific Prompts: Dataset-specific prefixes provide I-RAVEN layout and rule constraints, while Big-Bench Hard prefixes specify templates for temporal, logical, and tracking tasks.These prefixes precede the general developer prompt in the reasoning-structure stage.
B.2 System Architecture
Graphionale’s system architecture is presented as a prototype for generating visually augmented rationales, including an example for a Logical Deduction task.
- Graphionale’s system architecture is documented in Figure B3.
- The system includes an example of a visually augmented rationale for a BBH Logical Deduction (7 Objects) task.
C.1 Example Tasks from BIG-Bench Hard
This section presents example reasoning tasks and the prompt workflow used to convert argument maps into textual rationales.
- Example Tasks from BIG-Bench Hard: Figure C5 presents three example tasks from Big-Bench-Hard.
- Example Tasks from BIG-Bench Hard: Figure C6 presents three example tasks from I-RAVEN.
- Prompt Workflow: The workflow takes a JSON argument map as input for rationale generation.
- Prompt Workflow: The generated rationale uses headings, bullets, dividers, and explicit objection markers to preserve logical flow.