Towards Human-centered Proactive Conversational Agents

• Proactive conversational agents initiate actions toward long-term conversational goals rather than merely reacting to user requests.Examples include clarifying ambiguity, eliciting preferences, and directing open-domain conversations.Source: §1 · p. 1
• Without thoughtful design, initiative-taking systems may be perceived as intrusive, making human-centered design central to their effectiveness and acceptance.Source: §1 · p. 1
• The paper proposes a taxonomy of human-centered PCAs based on Intelligence, Adaptivity, and Civility.Source: §1 · p. 1, Figure 1 · p. 2
• It reinterprets prior PCA research and develops a research agenda across Task Formulation, Data Preparation, Model Learning, Evaluation, and System Deployment.Source: §1 · pp. 1–2

2.1 Key Dimensions

• Intelligence means anticipating task developments and planning strategically; Adaptivity means adjusting intervention timing and pacing to context; Civility means respecting personal, task, and ethical boundaries.Source: §2.1 · p. 2

2.2 Types of Proactive Conversational Agents

• The taxonomy categorizes PCAs into eight types according to high or low proficiency across the three dimensions, including Sage, Opponent, Boss, Cosseter, Listener, Airhead, Doggie, and Maniac.Source: §2.2 · pp. 2–3
• A Sage combines high Intelligence, Adaptivity, and Civility, whereas types such as Opponent and Cosseter illustrate how low Civility or Adaptivity can make capable systems intrusive or controlling.Source: §2.2 · p. 2
• Commercial conversational systems predominantly show foundational proactive Intelligence, while Adaptivity and Civility remain important areas for developing systems centered on human needs.Source: §2.2 · p. 3

2.3 Five Stages for PCA System Construction

• PCA construction proceeds through Task Formulation, Data Preparation, Model Learning, Evaluation, and System Deployment, with human-centered design intended throughout each stage.Source: §2.3 · p. 3

3 TASK FORMULATION

• Human-centered task formulation should integrate user emotions, preferences, values, and ethical standards alongside goal completion to foster trust and satisfaction.Source: §3 · pp. 3–4
• Mixed-initiative information seeking should combine clarification, out-of-scope handling, and extra information with prediction of when system initiative is appropriate.Source: §3 · p. 3
• Emotional support dialogues should move beyond mirroring feelings by selecting proactive interventions and timing initiative to address users’ concerns and emotional states.Source: §3 · pp. 3–4
• Pro-social negotiation should constrain tactics that humiliate or provoke users while promoting polite and empathetic interactions.Source: §3 · p. 4
• Personalized target-guided dialogue should pursue targets smoothly and align them with user interests and needs rather than assigning targets solely for system or seller objectives.Source: §3 · p. 4

4.1 Issues on Current Data Preparation Schemes

• Context-based datasets often fabricate user needs through truncated queries, assigned targets, imagined emotional problems, or predefined roles, limiting PCA Adaptivity.Source: §4.1.1 · p. 4, §4.1 · p. 4

4.1.2 Ethical Concerns.

• Crowdworker-annotated datasets generally show lower toxicity than datasets collected from raw counseling conversations, including HOPE and MI compared with ESConv.Source: §4.1.2 · p. 5
• Real-world dialogue collection raises privacy, toxicity, quality, and resource concerns, while crowdworker and generative-AI annotations introduce constrained patterns, shallow context understanding, and possible bias propagation.Source: §4.1.3 · p. 5, §4.1.2 · p. 5

4.2.2 Human-AI Collaborative Data Collection.

• The data-preparation discussion distinguishes real-world collection, crowdworker annotation, and generative-AI annotation as schemes with different limitations and trade-offs.Source: §4.1.3 · p. 5, Figure 3 · p. 6

4.2 New Perspectives on Data Preparation

• Human-centered collection should emulate real-world scenarios through authentic user data or annotators’ own needs and backgrounds, while acknowledging quality and ethical difficulties.Source: §4.2.1 · p. 5, §4.2 · p. 5

4.2.1 Reflecting Real Human Needs.

• Human-AI collaboration can combine human teachers’ nuanced understanding with LLM scalability and controllability, while expert guidance and AI reevaluation address annotation diversity and ethical concerns.Source: §4.2.2 · p. 5

5 MODEL LEARNING

• Human-centered PCA model learning organizes alignment approaches around Intelligence, Adaptivity, and Civility rather than planning capability alone.The paper discusses these dimensions through prompting, data augmentation, and human-feedback methods.Source: §5 · p. 6

5.1 Prompting by Human Instructions

• Prompting methods use human instructions and planning schemes such as CoT and multi-agent debate to improve proactive task solving.ProCoT can mitigate aggressive topic shifts, but its adaptivity benefits remain insufficiently understood and below desired levels.Source: §5.1 · p. 6
• Prompt-based methods remain limited because they struggle with long-term anticipation, transfer poorly to unseen scenarios, and cover only selected civility abilities.Manually designed prompts may improve honesty and thoughtfulness but are constrained to specific strategies such as Manners and Moral Integrity.Source: §5.1 · p. 6

5.2 Data Augmentation with Human Knowledge

• Human-knowledge data augmentation uses self-chat distillation and role-play simulation, but augmented data inherits planning limits and remains bounded by available human knowledge.Morality-related augmentation can support prosocial conversation, yet models may learn what to do without learning what to avoid.Source: §5.2 · p. 6

5.3 Learning from Human Feedback

• RLHF aligns language models with human preference signals, while simulated goal-oriented feedback can enhance PCA planning and anticipation.Its challenges include unreliable human supervision for future superintelligent systems, high costs for diverse feedback, and dependence on feedback quality.Source: §5.3 · pp. 6–7
• Table 3 evaluates prompt-based, supervised fine-tuning, and RL-based PCAs on ESConv across Intelligence, Adaptivity, and Civility.The analysis emphasizes that stronger Intelligence-related scores do not necessarily accompany better Adaptivity or Civility.Source: Table 3 · p. 7

6 EVALUATION

• The proposed framework evaluates PCAs across Intelligence, Adaptivity, and Civility rather than relying only on goal completion and response quality.The paper motivates this expansion because existing evaluations mainly overlook Adaptivity and Civility.Source: §6 · p. 7, §6.1 · p. 7

6.1.1 Evaluation Protocols for Adaptivity.

• Patience, Timing Sensitivity, and Self-awareness operationalize Adaptivity through semantic smoothness, turn-level satisfaction, and calibration error.The protocols use contextual semantic similarity, LLM-based satisfaction prediction, and Expected Calibration Error, respectively.Source: §6.1.1 · p. 7
• Civility is assessed through Boundary Respect, Moral Integrity, Trust and Safety, Manners, and Emotional Intelligence using Perspective API attributes and Emotional Intensity Relaxation.The first four abilities are scored through Identity Attack, Toxicity, Threat, and Insult, while Emotional Intelligence uses Emotional Intensity Relaxation.Source: §6.1.2 · p. 7

6.2 Empirical Analysis

• Methods perform diversely across the three dimensions: Alignedd-PM leads Intelligence but underperforms on some Adaptivity and Civility metrics, while ExTES improves Adaptivity over AugESC.Compared with standard RLHF, Alignedd-PM improves Adaptivity by incorporating diversified user preferences; ExTES benefits from human-designed strategies and counseling cases.Source: §6.2 · pp. 7–8

6.3 Prospects on Multidimensional Evaluation

• Higher Intelligence scores do not necessarily coincide with better Adaptivity or Civility, exposing limitations in current evaluation methods.The paper argues that multidimensional evaluation should be developed further, with human interactive judgments remaining most effective for Adaptivity and Civility despite scaling and standardization challenges.Source: §6.3 · p. 8, §6.3.1 · p. 8

6.3.1 Robust Evaluation Protocols.

• Evaluation gold standards should be customized to PCA types and social contexts rather than requiring high proficiency in all three dimensions.For example, Listener-type agents may not need Intelligence, while emergency scenarios can shift priorities from privacy toward physical safety.Source: §6.3.2 · p. 8

7 SYSTEM DEPLOYMENT

• Deployment should emphasize human behavior patterns and minimally intrusive system-initiated interactions rather than merely extending technological functionality.Language interfaces can be imprecise or inefficient for some tasks, so other modalities may improve user comfort and task efficiency.Source: §7 · p. 8, §7.1 · p. 8

7.1 Human-centered Designs of User Interface

• Conversational recommender systems can elicit preferences through language, Yes/No or multiple-choice questions, and comparison interfaces.Alternative interfaces can reduce the burden of interpreting natural-language preferences and let users convey preferences by tapping device icons.Source: §7.1 · p. 8

7.2 New Emphasis on Trust and Reliance

• Human-centered PCAs should foster appropriate reliance and trust by distinguishing whether users accept or reject behaviors according to the agent’s Adaptivity and Civility.Underreliance and undertrust can impede goal achievement, whereas overreliance and overtrust may pose risks to users.Source: §7.2 · pp. 8–9
• Explanations can reduce overreliance and overtrust through feature-based, example-based, path-based, and attribution-based forms.These explanation types respectively expose feature contributions, similar or representative examples, decision paths, and input or knowledge attributions.Source: §7.2.1 · p. 9

7.2.2 Reliability.

• Reliability disclosure communicates estimated system confidence, while self-calibration concerns how accurately the agent estimates and communicates that confidence.Showing confidence alongside a decision can significantly affect users’ trust and reliance.Source: §7.2.2 · p. 9
• User-controlled, on-demand suggestions can address underreliance and undertrust by letting people decide whether to view system-initiated content.The paper gives “See AI’s suggestion” as an example of exposing clarification questions or recommendations only after user request.Source: §7.2.3 · p. 9

8 CONCLUSIONS

• The paper proposes a taxonomy of Intelligence, Adaptivity, and Civility for interpreting PCA research from a human-centered perspective.It applies the taxonomy across Task Formulation, Data Preparation, Model Learning, Evaluation, and System Deployment.Source: §8 · p. 9
• The authors use limitations of current PCA research to envision a future agenda for achieving human-centered proactive conversational agents.The conclusion emphasizes maintaining human-centered systems as PCAs advance toward superintelligent AI so they continue serving human interests.Source: §8 · p. 9