Rethinking Group Recommender Systems in the Era of Generative AI: From One-Shot Recommendations to Agentic Group Decision Support

Technical Summary: Rethinking Group Recommender Systems in the Era of Generative AI

Problem Statement
Despite over twenty-five years of research into Group Recommender Systems (GRS), the field suffers from a significant gap between academic algorithmic development and real-world adoption. Current literature heavily emphasizes preference aggregation algorithms and offline evaluation metrics, often neglecting the complex, interactive nature of human group decision-making. Existing systems frequently assume a "one-shot" interaction model where users explicitly state preferences, which are aggregated into a ranked list. This approach fails to align with how groups naturally communicate (e.g., via messenger apps), negotiate, and reach consensus. Furthermore, the limited number of deployed real-world GRS prototypes suggests that current assumptions regarding inter-group communication and decision support may be misaligned with user needs.

Methodology and Proposed Framework
The paper does not present a new empirical dataset or a specific implemented system with experimental results. Instead, it proposes a forward-looking conceptual framework for "Agentic Group Decision Support." The methodology involves a critical reflection on existing GRS literature, human decision-making theories, and conversational decision-support systems, integrated with the emerging capabilities of Large Language Models (LLMs) and Generative AI.

The core of the proposal is a shift from reactive, list-based recommendation to an agentic, chat-based approach. The envisioned system operates within existing instant messaging environments (e.g., WhatsApp, Telegram) where a group of users and an AI agent interact. The agent is designed not merely as a recommendation engine but as a facilitator of the decision-making process.

The technical architecture draws upon recent frameworks for LLM-based autonomous agents, specifically adapting the Profile, Memory, Planning, and Action modules:

• Profile: The agent constructs in-depth representations of users based on observed behaviors, social traits, and preferences inferred from chat interactions, rather than relying solely on explicit ratings.
• Memory: Acts as a "contextual brain," storing interaction history, emotional states, and group dynamics to maintain a shared mental model of the decision state.
• Planning: Determines the agent's strategy, deciding when to intervene proactively (e.g., summarizing discussions, soliciting silent members) versus reacting to user prompts. It manages multi-step action plans to guide the group toward a decision.
• Action: Executes tasks such as generating recommendations, retrieving external information (e.g., weather, opening hours), summarizing conversations, and mediating conflicts.

The system supports various decision-making principles identified in literature, such as attribute-based choice and socially-based choice, by providing meta-level advice, visualizing consensus/disagreement, and de-escalating conflicts.

Key Contributions

1. Paradigm Shift: The paper argues for a fundamental re-orientation of GRS research from "preference aggregation" to "decision process support." It posits that the primary role of a GRS should be facilitating communication and consensus among group members rather than acting as a normative judge.
2. Agentic Integration: It outlines how Generative AI and LLMs can enable "agentic" behaviors in GRS, allowing the system to monitor group dynamics, identify silent members, manage turn-taking, and provide real-time summaries of complex discussions—capabilities that were technically infeasible in the pre-LLM era.
3. Configurable Framework: The authors propose a "class of systems" approach rather than a single solution. The agent's behavior (proactive vs. reactive, moderator vs. participant) is configurable based on the specific use case and group context.
4. Interdisciplinary Synthesis: The work synthesizes insights from recommender systems, human-computer interaction (HCI), social psychology (group dynamics), and group decision support systems (GDSS) to define a more holistic design space.

Results and Claims
As a perspective paper, the authors do not report experimental results, user studies, or performance benchmarks. The "results" are theoretical and conceptual:

• The paper demonstrates that current GRS limitations (low adoption, lack of natural interaction) stem from a misalignment with real-world decision processes.
• It argues that LLMs possess the necessary capabilities (zero-shot/few-shot recommendation, text summarization, intent recognition, and reasoning) to implement the proposed agentic framework.
• It identifies that while LLMs offer new opportunities, they also introduce specific challenges, including hallucinations, non-deterministic behavior, the difficulty of modeling self-awareness in group settings, and the complexity of prompt engineering for multi-module coordination.

Significance
The paper claims significance as a call to action for the research community to move beyond algorithmic optimization of preference aggregation. It suggests that the integration of Generative AI offers a viable path to solving the "adoption problem" of GRS by embedding recommendation support into the natural communication channels groups already use. By shifting the focus to the process of decision-making and leveraging LLMs as active facilitators, the authors believe the field can achieve broader practical adoption. The paper emphasizes that future research must be interdisciplinary, involving social scientists and HCI experts, and must evolve evaluation methodologies beyond offline metrics to include human-centric assessments of decision quality, fairness, and process satisfaction.