PaperlayerView exact revision on arXiv

Source-linked AI summary

Feedback by Design: Understanding and Overcoming User Feedback Barriers in Conversational Agents

Nikhil Sharma, Zheng Zhang, Daniel Lee, Namita Krishnan, Guang-Jie Ren, Ziang Xiao, Yunyao Li

arXiv:2602.01405v1cs.HC

TL;DR

Users rarely provide the clear, structured feedback that conversational agents need, despite its importance for interaction and model improvement. The paper combines formative and within-subject studies to identify Feedback Barriers and test model-agnostic interface scaffolds. FeedbackGPT enabled higher-quality and more frequent feedback, especially more goal-referenced feedback, while the study’s sample and scaffold configurations limit generalizability.

  • Problem

    High-quality feedback is important for human-AI interaction, but users often provide it infrequently and in fragmented, low-quality forms.

  • Method

    Two studies identified four Feedback Barriers through formative interviews and evaluated FeedbackGPT’s model- and task-agnostic scaffolds in a within-subject study.

  • Results

    FeedbackGPT enabled users to provide significantly more goal-referenced, actionable, and progressive feedback, with relevant feedback increasing from 32.49% to 58.43%.

  • Takeaways & Limitations

    Interface scaffolds that minimize Feedback Barriers can support more frequent and higher-quality feedback without changing model weights or core model behavior.

  • Takeaways & Limitations

    The participant sample was not representative and the study had selection bias, while the evaluated scaffolds were only one non-exhaustive instantiation of the desiderata.

Abstract

from arXiv · show

High-quality feedback is essential for effective human-AI interaction. It bridges knowledge gaps, corrects digressions, and shapes system behavior; both during interaction and throughout model development. Yet despite its importance, human feedback to AI is often infrequent and low quality. This gap motivates a critical examination of human feedback during interactions with AIs. To understand and overcome the challenges preventing users from giving high-quality feedback, we conducted two studies examining feedback dynamics between humans and conversational agents (CAs). Our formative study, through the lens of Grice's maxims, identified four Feedback Barriers -- Common Ground, Verifiability, Communication, and Informativeness -- that prevent high-quality feedback by users. Building on these findings, we derive three design desiderata and show that systems incorporating scaffolds aligned with these desiderata enabled users to provide higher-quality feedback. Finally, we detail a call for action to the broader AI community for advances in Large Language Models capabilities to overcome Feedback Barriers.

1 Introduction

High-quality feedback is important for aligning conversational agents with users’ evolving goals, yet users often provide feedback infrequently and ambiguously. The paper studies these barriers and introduces FeedbackGPT scaffolds designed to elicit more frequent and higher-quality feedback.

  • Human feedback helps conversational agents remain aligned with users’ evolving intent and requirements during multi-turn, goal-oriented tasks.
  • Only 38,992 of 1.02 million ChatGPT conversations in WildFeedback contained feedback, and the feedback was often ambiguous or incomplete.
  • The paper asks what barriers limit feedback with LLM-powered conversational agents and how scaffolds can support sustained, high-quality feedback.
  • A formative study with 16 participants examined breakdowns and feedback practices across diverse goal-driven CA tasks through the lens of Grice’s maxims.
  • A second within-subject study with 20 participants evaluated FeedbackGPT, which incorporated six model-agnostic scaffolds targeting barriers identified in the formative study.
  • Mixed-methods analysis found that scaffolds minimizing feedback barriers enabled users to provide higher-quality and more frequent feedback.
  • The paper contributes empirical insights into Feedback Barriers, FeedbackGPT, and practical recommendations for eliciting high-quality feedback.

2 Related Works

Prior work treats feedback as a central mechanism for steering interactive systems, but LLM-powered conversational agents create distinctive challenges in open-ended, evolving multi-turn collaboration. This paper addresses limited attention to the interactional work and barriers involved in giving feedback.

  • Feedback research spans explicit ratings or corrections, implicit behavioral signals, mixed-initiative systems, interactive machine learning, and explanatory debugging.
  • Conversational recommender research shows that interface choices shape when and how users provide corrections and refine suggestions over time.
  • LLM-powered agents differ from traditional interactive systems because they support open-ended multi-turn tasks with evolving goals, strategies, and evaluation criteria.
  • Existing ML-centric work finds that usable feedback data are scarce, fragmentary, and rarely capture the rich structure needed for nuanced alignment.
  • Recent HCI work studies LLM agents as mixed-initiative collaborators, real-world usage traces, and interaction experiences beyond static task accuracy.
  • The paper shifts attention toward turn-level feedback acts, conversational breakdowns, and interface scaffolds that make feedback easier to express, interpret, and incorporate.
  • Feedback is generally timely, specific, and goal-referenced when designed to help people close the gap between current performance and a desired standard.

3 Formative Study

The formative study found that users commonly encounter context drift, unreliable outputs, and instruction-following problems, then respond with short-lived, reactive repair strategies. Through Grice’s maxims, the study characterizes reciprocal breakdowns that limit feedback quality.

  • What are the common breakdowns while interacting with CAs?: Context loss was the dominant interaction failure, reported by 10 of 16 participants, as agents gradually drifted from the original task goal.
  • What are the common breakdowns while interacting with CAs?: Participants also encountered fabricated or unreliable content, along with instruction-following failures and sycophantic agreement.
  • How do users give feedback?: Users repaired derailments by rephrasing or repeating requests, refining inputs, decomposing tasks, restarting conversations, switching tools, or completing tasks themselves.
  • How do users give feedback?: Feedback was usually reactive, local to the latest turn, and short-lived rather than rich, structured, ongoing guidance.
  • How do users give feedback?: High-quality feedback was assessed as goal-referenced, actionable, articulated, and progressive, but users often supplied granular corrections that signaled dissatisfaction without process guidance.
  • How do users give feedback?: Users’ comments were almost entirely task-level, nudging specific outputs rather than providing process-level guidance for calibrating collaboration over time.
  • Maxim of Relation: CAs shift the conversational goal and users: Violations of Relation included model goal drift and user withdrawal, while violations of Quality included hallucinations, verification burdens, and relaxed evidential standards.
  • Maxim of Quality: CAs hallucinate and users propagate unver-: Violations of Manner made responses difficult to interpret because models often failed to clarify or adjust and lacked human paralinguistic cues.

4 Designing systems to resolve the Feedback barriers

The paper translates four Feedback Barriers into three design desiderata addressing shared context, reciprocal interaction, and verifiable outcomes. FeedbackGPT operationalizes these desiderata through interface scaffolds that lower the effort of giving, refining, and evaluating feedback.

  • The four Feedback Barriers disrupt feedback dissemination, engagement with feedback, and evaluation of feedback outcomes.
  • The design desiderata aim to reduce interaction costs that prevent users from giving high-quality feedback.
  • Foster a Persistent and Legible Shared Frame of Reference: The first desideratum requires a persistent, visible, editable shared understanding of task goals and constraints rather than reliance on transient chat history.
  • Design for Proactive and Low-Friction Interaction: The second desideratum uses proactive, mixed-initiative clarification and structured actions to lower the effort of expressing and refining feedback.
  • Provide Transparent and Verifiable Reasoning: The third desideratum makes reasoning and feedback effects transparent and auditable, for example through visual diffs that support low-cost evaluation.
  • Foster a Persistent and Legible Shared Frame of Reference: Inline Comments & Highlights anchor feedback to output spans, preserve it in conversation state, and pass structured constraints to the next input.
  • Foster a Persistent and Legible Shared Frame of Reference: Undo and Redo let users explore alternative feedback paths without discarding accumulated conversation context or restarting the interaction.
  • Design for Proactive and Low-Friction Interaction: Feedback Huddle asks targeted questions, Quick Actions provide reusable templates, and Feedback Evaluation offers real-time suggestions for specificity, structure, and clarity.

5 Study 2: How do affordances that resolve Feedback Barriers change user feedback?

Study 2 evaluated FeedbackGPT in a within-subject comparison with ChatGPT, using two goal-oriented co-writing tasks and measures of feedback quality, user experience, and feedback barriers.

  • Study design: 20 participants completed two co-writing tasks, using ChatGPT for one task and FeedbackGPT for the other.Each session lasted approximately one hour, and participants spent at least 20 minutes interacting with each assigned system.
  • Study procedure: The procedure included pre-task questions, two system-assigned tasks, task-specific surveys, a post-task survey, and an interview.Participants selected topics from three options and received the system not used in their first task for the second task.
  • Measures: Participants rated output satisfaction, conversation satisfaction, and perceived productivity on five-point Likert scales after each task.Post-task questions also assessed ease of providing feedback, comparative system preferences, and experiences of each feedback barrier.
  • Task design: Co-writing tasks were chosen because they required multiple turns and user feedback to bridge a knowledge gap with the model.The tasks involved users’ personal experiences and avoided specialized interface requirements.
  • Measures: Feedback quality was measured through goal-referenced, actionable, articulated, and progressive dimensions, alongside subjective experience and barrier measures.The study also collected pre-task attitudes and demographics for contextualization.
  • Analysis: A feedback turn was defined as a post-response user message containing evaluative, corrective, or guiding content intended to shape the system’s subsequent output.Researchers extracted feedback turns and the preceding model response from both system conditions.

6 Results

FeedbackGPT produced more goal-referenced, specific, and voluminous feedback than ChatGPT, while objective clarity did not significantly improve. Users associated the system with easier guidance and greater control, but reported increased effort.

  • Barrier exposure: FeedbackGPT exposed comparable model-side barrier rates across conditions, supporting a fair comparison of scaffold effects.GEE models found no significant system effect for Relation, Verifiability, Disambiguation, or Informativeness.
  • Goal-referenced: FeedbackGPT participants’ goal-referenced feedback rose from 32.49% ± 7.07 with ChatGPT to 58.43% ± 6.56, a +25.94% gain.The difference was significant (n=20, W=40.5, p< .05).
  • Mechanisms: Participants linked FeedbackGPT’s gains to anchored comments, comparisons, inline controls, locus-specific annotation, and prompts for additional details.These features helped preserve conversation state, localize edits, and elicit more detailed inputs.
  • Actionable: Specificity increased from 65.86% ± 4.58 with ChatGPT to 90.62% ± 2.60 with FeedbackGPT, a +24.76 increase.The difference was significant (n=20, W=1.0, p< 0.01).
  • Articulate: Clarity did not differ significantly: 77.59% ± 6.59 with ChatGPT versus 84.85% ± 4.77 with FeedbackGPT, Δ = +7.26.Participants perceived improved communication, but vague inputs persisted and conversation-quality ratings did not significantly improve.
  • Progressive: Feedback length increased from 242.33 ± 35.24 characters per turn with ChatGPT to 585.06 ± 68.55 with FeedbackGPT.The difference was significant (n=20, W=32.0, p< 0.01), and users rated FeedbackGPT higher for Informativeness.
  • User experience: The richer feedback workflow increased cognitive load even as participants felt more productive, in control, and better able to guide the AI.Participants also rated feedback as easier with FeedbackGPT (x̄= 3.69, p< 0.01) and reported improved ability to guide the AI (x̄= 3.63, p< 0.01).

7 Discussion

The studies identify four Feedback Barriers and show that lightweight scaffolds can make feedback more goal-referenced, actionable, and progressive, while increasing users’ cognitive load. The discussion extends these findings into design principles and model-level research priorities.

  • The studies identify Common Ground, Verifiability, Communication, and Informativeness as Feedback Barriers arising from model limitations and user-side cognitive constraints.
  • From Monologue to Dialogue: Interpreting the Impact of Scaffolding: FeedbackGPT’s scaffolds shifted interaction from brittle, command-based exchanges toward more collaborative dialogue and enabled more precise, actionable feedback.
  • From Monologue to Dialogue: Interpreting the Impact of Scaffolding: 9/20 users still felt the burden of providing the right feedback remained on them, indicating that effective scaffolding can increase perceived cognitive load.
  • From Monologue to Dialogue: Interpreting the Impact of Scaffolding: Inline comments anchored feedback to specific locations, helping users express intent precisely and reducing lengthy, decontextualized explanations.
  • Design Implications for Truly Collaborative Agents: The proposed design principles call for proactive mixed initiative, reciprocal mutual grounding, and clearer evidence that feedback is used beyond the immediate turn.
  • A call for action for the broader Artificial Intelligence community: Overcoming Feedback Barriers also requires model advances in context-robust memory, calibration, principled abstention, and reliable inline citation.
  • A call for action for the broader Artificial Intelligence community: Scaffolded interfaces could produce structured preference data, including span-level highlights, comments, huddle transcripts, and prompt-evaluation scores, for future preference-learning pipelines.

8 Limitations

The study’s findings are constrained by participant selection, task scope, scaffold coverage, and the short-term design of the evaluation. These limitations leave interface-specific, interaction-combination, and longitudinal effects for future work.

  • The participant sample was not representative, with gender and education imbalances and selection bias among volunteers completing the cognitively demanding second study.
  • Because the study focused on co-writing, other task-specific interfaces may elicit different user behaviors.
  • The scaffolds represent one instantiation of the design desiderata, and the study did not examine how different actions interact to elicit feedback behaviors.
  • The short-term study cannot predict how feedback behaviors will evolve as users become accustomed to the interfaces and model behaviors.

9 Conclusion

The paper shows that high-quality feedback depends on both model behavior and interaction design. Lightweight scaffolds addressing four feedback barriers increased the relevance, specificity, and volume of user feedback.

  • Four Feedback Barriers—Common Ground, Verifiability, Communication, and Informativeness—arise from reciprocal user and model breakdowns.
  • Lightweight, model-agnostic scaffolds operationalizing persistent shared reference, reciprocal interaction, and verifiable reasoning increased feedback relevance, specificity, and volume.
  • The authors call for advances in memory, calibration, and multi-turn training alongside interaction scaffolds that sustain cooperative feedback norms.
  • The broader goal is to make feedback cognitively affordable, socially reciprocal, and visibly impactful, supporting more trustworthy, adaptive, and socially aligned AI systems.

A.2 Detailed task descriptions

Participants were instructed to apply for a $250,000 STEM Fellowship focused on transformative artificial-intelligence visions in their fields.

  • The task asks applicants to propose bold, creative uses of artificial intelligence in transformative ways within their respective fields.

A.2.1 Task 1: $250K STEM Fellowship Application.

The STEM Fellowship application requires a 500–1000-word essay combining personal narrative, qualifications, an innovative AI vision, and fellowship impact. Applicants must also deliver a talk for the essay’s first minute.

  • 500–1000 words is mandatory, and applications outside this range are disqualified.
  • Applicants must share their background, values, motivations, strengths, experiences, and perspectives.
  • The essay must present an innovative AI project or initiative and explain its personal and wider positive social impact.
  • Applicants must deliver a talk covering the first minute of the drafted essay.

A.2.2 Task 2: $50K Neighborhood Improvement Grant.

The supplied task materials describe a neighborhood-improvement grant and related pitch requirements, including a 500–1000-word limit, a five-minute presentation, and a one-minute audio submission. They also include a separate $10K competition prompt and a participant-use table caption.

  • A.2.2 Task 2: $50K Neighborhood Improvement Grant.: Applicants must describe a neighborhood problem, concrete solution, qualifications, grant use, sustainability, and community impact.
  • A.2.2 Task 2: $50K Neighborhood Improvement Grant.: The neighborhood grant requires 500–1000 words and a talk covering the pitch’s first minute.
  • A.2.2 Task 2: $50K Neighborhood Improvement Grant.: The separate $10K competition requires a five-minute personal pitch of approximately 500–1000 words, plus a transcript and one-minute audio recording.
  • A.2.2 Task 2: $50K Neighborhood Improvement Grant.: The $10K pitch should explain the applicant’s work, distinctive expertise, leadership readiness, requested next step, and expected impact.
  • A.2.2 Task 2: $50K Neighborhood Improvement Grant.: Table 1 is captioned as reporting participant demographics and AI-use-case expertise, identifying P12–P16 as inexperienced users.
Loading 2602.01405v1…