Graphionale: How Graph Visualizations of LLM Rationales Affect Human Decision Making

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.

Introduction. Recent advances in Large Language Models (LLMs) have significantly improved their reasoning capabilities. With these developments, LLMs are increasingly used as conversational assistants for complex, knowledge-intensive questions. To improve transparency and support human decision making, LLMs provide detailed rationales describing how an answer is derived [4, 34, 62, 65, 66]. However, these rationales are traditionally presented as long blocks of text. Even when organized into sections, text remains inherently linear and cannot intuitively represent relational structures among reasoning steps, such as branching, aggregation, or elimination. As a result, users must parse sequential sentences to reconstruct the underlying logic, making it difficult to quickly grasp the overall reasoning or evaluate the model’s conclusions [23, 37]. In contrast, humans often externalize complex explanations into diagrams to support understanding. Students draw concept maps to visualize relationships between ideas [35, 44].

Discussion / Conclusion. Our results show that graphical rationales did not help uniformly: they improved trust calibration for verbal reasoning yet impaired it for visual reasoning, while subjective ratings showed the opposite pattern in each domain. We discuss the implications of these findings in this section. The reversal in effectiveness across task modalities can be explained by the complementarity of representations [25, 49]. A diagram helps when its structure matches the structure of the problem. It allows problem-related operations to be done perceptually rather than symbolically, and reduces the working memory load [25]. In verbal reasoning, the argument graph achieves this match— each node maps to a logical inference and edges encode dependencies that participants can inspect sequentially. But visual reasoning operates under different representational demands.