"What Are You Doing?": Effects of Intermediate Feedback from Agentic LLM In-Car Assistants During Multi-Step Processing

Paper Title

'What Are You Doing?': Effects of Intermediate Feedback from Agentic LLM In-Car Assistants During Multi-Step Processing

Publication Info

• Topic area: Feedback design for agentic AI systems in dual-task contexts, specifically in-car assistants.
• Keywords: Agentic AI, in-car assistants, feedback timing, LLM-based systems, trust, user experience, cognitive load, adaptive feedback, dual-task interaction, human-AI interaction.

Background and Problem

• Problem / challenge: Current agentic AI systems lack standardized principles for feedback timing and verbosity during multi-step tasks. This creates challenges in balancing transparency, trust, and cognitive load, especially in dual-task contexts like driving.
• Significance: Poorly designed feedback can lead to user frustration, cognitive overload, or reduced trust, while insufficient communication creates "ambiguous silence," undermining user confidence in system performance.
• Motivation and related work: Existing systems vary widely in feedback strategies, from silent operation to verbose updates, but lack empirical grounding. Prior research highlights the importance of transparency, responsiveness, and trust in human-AI interaction, but it remains unclear how these principles apply to agentic systems with extended processing times.

Solution

• Proposed approach: Investigate the effects of intermediate feedback timing and verbosity in an LLM-based in-car assistant using a mixed-methods study.
• Novelty:
  1. Empirical evidence showing intermediate feedback improves perceived speed, trust, and user experience while reducing task load.
  2. Identification of user preferences for adaptive feedback that evolves based on trust, task stakes, and situational context.
  3. Design implications for feedback timing and verbosity in dual-task contexts, with potential applicability to other domains.
• Procedure and key techniques:
  • Conducted a controlled user study (N=45) in a car simulation environment.
  • Compared two feedback strategies: No Intermediate (NI) feedback vs. Planning & Results (PR) feedback.
  • Measured perceived speed, task load, user experience, and trust across varying task durations and interaction contexts (stationary vs. driving).
  • Complemented quantitative results with qualitative interviews to explore preferences for adaptive feedback.

Results

• Concrete findings:
  • Intermediate feedback (PR) significantly improved perceived speed (dz = 1.01), user experience (dz = 0.54), and trust (dz = 0.38), while reducing task load (dz = −0.26).
  • PR feedback buffered the negative impact of longer task durations on perceived speed.
  • Task load reduction was primarily driven by lower frustration levels.
• Advantage over baselines:
  • PR feedback outperformed NI feedback across all metrics, including responsiveness, cognitive load, and trust.
  • Stepwise updates were perceived as cognitively lighter than dense final responses.
• Experiments / evaluation:
  • Quantitative study: 2×2×2 factorial design with independent variables (feedback timing, task duration, interaction context) and dependent variables (perceived speed, task load, user experience, trust).
  • Qualitative study: Thematic analysis of interviews revealed preferences for adaptive feedback based on trust, task stakes, and situational context.
• Limitations and future work:
  • Study conducted in a simulated driving environment, limiting real-world generalizability.
  • Feedback adaptation was assessed qualitatively rather than through longitudinal behavioral data.
  • Future work should explore adaptive feedback policies, multimodal feedback combinations, and domain-specific constraints.

Summary

This study demonstrates that intermediate feedback significantly enhances perceived speed, trust, and user experience while reducing task load in LLM-based in-car assistants. Users prefer adaptive feedback that starts with high transparency to build trust, then reduces verbosity as reliability is demonstrated, with situational adjustments for task stakes and context. These findings inform design principles for agentic systems in dual-task contexts, with potential applicability to other domains requiring extended multi-step processing.