[Paper] FeedAIde: Guiding App Users to Submit Rich Feedback Reports by Asking Context-Aware Follow-Up Questions

Source: arXiv - 2603.04244v1

Overview

Mobile apps thrive on user feedback, but the reports users submit are often vague and missing crucial context, forcing developers into lengthy back‑and‑forth clarification. FeedAIde tackles this problem by embedding a multimodal large language model (LLM) directly into the app’s feedback flow, automatically asking smart, context‑aware follow‑up questions (e.g., “What were you doing when this screenshot was taken?”). The result is richer, more actionable bug reports and feature requests without burdening users.

Key Contributions

• Context‑aware feedback loop: Captures the current UI screenshot and feeds it to a multimodal LLM, which generates adaptive follow‑up questions tailored to the observed problem.
• iOS‑native framework: Provides a reusable SDK that developers can drop into any iOS app to enable AI‑driven feedback collection.
• Empirical validation: Field study with a gym‑app’s real users showed higher perceived ease‑of‑use and usefulness compared with a traditional static feedback form.
• Quality boost for developers: Independent expert evaluation of 54 reports revealed significant improvements in completeness and relevance, especially for bug reports.
• Design guidelines: Offers practical recommendations for integrating GenAI‑powered questioning without overwhelming users.

Methodology

1. Data capture: When a user initiates feedback, the framework automatically grabs a screenshot and any available app state (e.g., current view hierarchy).
2. Prompt engineering: The screenshot and a brief user description are sent to a multimodal LLM (e.g., GPT‑4V). The prompt instructs the model to infer missing details and suggest concise follow‑up questions.
3. Interactive refinement: The app presents the generated questions one‑by‑one. Users answer, and each answer is fed back to the model to refine subsequent questions, converging on a complete report.
4. Report assembly: All user inputs, the original screenshot, and the conversation transcript are compiled into a structured feedback artifact (JSON + markdown) that developers can ingest directly into issue‑tracking tools.
5. Evaluation:
   • User study: 30 participants used the gym app with either FeedAIde or the native feedback form; post‑task surveys measured perceived ease, usefulness, and satisfaction.
   • Expert review: Two industry practitioners rated the resulting reports on completeness, reproducibility, and relevance using a 5‑point rubric.

Results & Findings

Key takeaways:

• Users felt the AI‑driven questions guided them without being intrusive.
• Developers received reports that already contained the “who, what, where, when” details they normally have to chase down.
• The multimodal aspect (screenshot‑aware questioning) was crucial; without visual context, the LLM’s follow‑ups were far less precise.

Practical Implications

• Faster triage: Development teams can cut down on clarification cycles, accelerating bug fixing and feature planning.
• Higher user satisfaction: Users see their feedback being taken seriously because the app helps them articulate the problem, reducing frustration.
• Scalable support: Companies can deploy the SDK across many apps, letting GenAI handle the first‑line “information gathering” step that traditionally required human support staff.
• Integration with existing pipelines: The structured JSON output can be hooked into Jira, GitHub Issues, or custom analytics dashboards, enabling automated tagging or priority assignment.
• Reduced QA overhead: More complete reports mean QA can reproduce issues earlier, decreasing the need for extensive exploratory testing.

Limitations & Future Work

• Platform scope: The current implementation is iOS‑only; Android and cross‑platform frameworks need separate adapters.
• Model dependency: Quality of follow‑up questions hinges on the underlying multimodal LLM; cost and latency could be concerns for large‑scale deployments.
• Privacy considerations: Automatic screenshot capture raises data‑privacy questions; future versions must incorporate on‑device inference or explicit user consent flows.
• User fatigue: While the study reported low fatigue, longer sessions or complex issues might lead to too many questions; adaptive stopping criteria are a promising research direction.
• Generalizability: The evaluation was limited to a single gym app; broader field studies across diverse domains (e.g., finance, gaming) are needed to confirm robustness.

FeedAIde demonstrates that embedding context‑aware generative AI directly into the feedback loop can bridge the long‑standing gap between what users say and what developers need—turning noisy, incomplete reports into actionable intelligence.

Authors

• Ali Ebrahimi Pourasad
• Meyssam Saghiri
• Walid Maalej

Paper Information

• arXiv ID: 2603.04244v1
• Categories: cs.SE, cs.AI, cs.HC
• Published: March 4, 2026
• PDF: Download PDF