Do visual rationales help or hurt how people calibrate trust?

Graphionale turns a linear LLM rationale into an interactive, multi-level graph that makes conclusions, premises, support and objections explicit, and then compresses each statement into entities and relations. The paper tests it in a large online study (N=204) that varies task modality (verbal vs. visual reasoning), rationale format (textual vs. graphical) and question difficulty (easy vs. hard). The headline result is that graphical rationales "did not help uniformly." For verbal reasoning they improved trust calibration, yet participants found them more cognitively demanding and less satisfying. For visual reasoning they impaired calibration, yet participants found them more engaging and helpful. In each domain the subjective ratings ran opposite to the calibration outcome.

The paper's explanation is the complementarity of representations: "a diagram helps when its structure matches the structure of the problem." A matched diagram lets problem-related operations be done perceptually rather than symbolically and reduces working-memory load. In verbal reasoning the argument graph achieves that match, because each node maps to a logical inference and edges encode dependencies participants can inspect sequentially. Visual reasoning "operates under different representational demands," so the same graph no longer fits the problem. The motivating complaint is that text is inherently linear and cannot show branching, aggregation or elimination, which forces readers to reconstruct the logic from sequential sentences.

This qualifies the argument that structure alone makes machine reasoning legible. Can formal argumentation make AI decisions truly contestable? describes structure as an explanatory property of the decision. Graphionale locates the effect at the reader: an explicit argument structure can help or hurt a human's calibration depending on the task. It also echoes When does explicit reasoning actually help model performance?, where the fit between format and task structure decides whether explicit reasoning helps the model. Here the same principle governs whether a rationale format helps the person reading it. The graphs organize the rationale text; they do not address whether that text tracks what the model computed, which Do language models actually use their reasoning steps? treats as a separate problem.

The excerpt is silent on how trust calibration was measured, which tasks counted as verbal or visual, and how large the effects were. It reports no difficulty results. The explanation for the visual-reasoning reversal is cut off after "different representational demands," and the complementarity account is offered as an explanation, not as something the study tested. What follows at this strength is narrow: a rationale interface should not be evaluated on satisfaction ratings alone, because in this study the format people liked less on verbal tasks was the one that calibrated their trust better, and the reverse held for visual tasks.