Sensemaking in Multi-Agent LLM Interfaces: How Users Interpret Transparency and Trustworthiness Cues

Paper Title

Sensemaking in Multi-Agent LLM Interfaces: How Users Interpret Transparency and Trustworthiness Cues

Publication Info

• Topic area: User interpretation of transparency and trustworthiness in multi-agent Large Language Model (LLM) interfaces.
• Keywords: Multi-agent systems, transparency, trustworthiness, LLM interfaces, user mental models, epistemic cues, trust calibration, human-AI interaction, design-led study, interface design.

Background and Problem

• Problem / challenge: Current transparency mechanisms in AI systems often fail to align with how users interpret and engage with outputs of generative LLMs, especially in multi-agent settings. There is limited understanding of how users perceive and evaluate transparency and trustworthiness in such systems.
• Significance: Understanding user needs for transparency in multi-agent systems is crucial for designing interfaces that foster appropriate trust and support effective decision-making.
• Motivation and related work: Previous research has explored trust calibration in classical AI systems and initial studies on multi-agent LLMs, but these efforts remain scattered and lack a principled understanding of how users interpret multi-agent reasoning and transparency cues. This paper addresses this gap by focusing on user mental models and preferences for multi-agent transparency.

Solution

• Proposed approach: A design-led, qualitative, comparative structured observation study using five interface variants of multi-agent LLMs to explore user perceptions of transparency and trustworthiness.
• Novelty:
  1. A systematic exploration of how users interpret multi-agent reasoning and transparency cues.
  2. A reconceptualization of transparency as a context-sensitive sufficiency judgment rather than a volume dial.
  3. Identification of design tensions between visibility, interpretability, and cognitive effort.
  4. Introduction of progressive, on-demand transparency as a design strategy.
• Procedure and key techniques:
  1. Literature review to define a design space for multi-agent transparency, identifying seven design dimensions.
  2. Development of five interface variants (V1–V5) operationalizing different transparency configurations.
  3. In-person lab study with 12 participants, using think-aloud protocols, card-sorting activities, and semi-structured interviews.
  4. Analysis of user interactions with interfaces across two task types: information-seeking and logical reasoning.

Results

• Concrete findings:
  • Users interpreted epistemic signals (e.g., disagreement, critique, consensus) as key cues for trustworthiness.
  • Participants preferred a "Goldilocks" level of transparency, balancing informational value and cognitive effort.
  • Task complexity and user expertise influenced transparency preferences, with simpler tasks requiring less visibility.
  • Progressive, on-demand transparency was widely desired to manage cognitive workload.
• Advantage over baselines:
  • Interfaces with agent-level rationales (e.g., V3) were perceived as more trustworthy and helpful compared to opaque designs (e.g., V1).
  • Explicit critique (V4) and debate (V5) enhanced trust by showing the system "checking itself."
• Experiments / evaluation:
  • Participants interacted with five interface variants across two tasks (information-seeking and reasoning).
  • Evaluations included perceived transparency, helpfulness, and reliability using card-sorting activities.
  • Data were analyzed using thematic analysis to identify user mental models and transparency preferences.
• Limitations and future work:
  • Study tasks were predefined and bounded, limiting exploration of self-directed use cases.
  • Interface configurations tested represent only a subset of the broader design space.
  • Future work should examine transparency preferences in scenarios with agent errors and test progressive disclosure mechanisms at scale.

Summary

This study explores how users interpret transparency and trustworthiness in multi-agent LLM interfaces. By analyzing user interactions with five interface variants, the authors identify key epistemic cues (e.g., disagreement, critique, consensus) that shape trust perceptions. Transparency is reconceptualized as a context-sensitive sufficiency judgment, with users preferring progressive, on-demand transparency to balance cognitive effort and informational value. Task complexity, user expertise, and dispositional trust influence transparency needs. These findings provide actionable insights for designing trustworthy, human-centered multi-agent AI systems and highlight the need for future work on adaptive transparency mechanisms.