Beyond Serendipity: From Exposing the Unknown to Fostering Engagement through Peer Recommendation

Abstract: Serendipity-oriented recommender systems expose users to unfamiliar items to counter filter bubbles, yet mere exposure does not ensure that users will understand or appreciate the content they encounter. We propose Peer Recommendation, a framework in which a user and an AI agent (Peer) with distinct preferences collaboratively explore unfamiliar content. Unlike conventional conversational recommender systems where the user is a passive recipient, our framework positions the user as both a recommender and a recipient: the user and the Peer mutually recommend songs to each other through chat-based dialogue, collaboratively building a shared playlist. In an exploratory within-subjects experiment (N=14), we compared three conditions: (1) a Close Peer, (2) a Distant Peer, and (3) a baseline agent without an explicit preference profile. The Close Peer significantly increased users' interest expansion and perceived value of the activity compared to the baseline, with medium-to-large effect sizes. The Distant Peer showed no significant difference at the aggregate level; however, qualitative analysis revealed varied responses, with some participants strongly preferring the Distant Peer. These findings suggest that the "otherness" of a recommendation partner is essential for moving beyond mere exposure toward genuine engagement, and that the appropriate degree of preference distance may vary and need to be adapted to individual users.

• The paper introduces a peer recommendation system that uses dialogue-based mutual recommendations to enhance user engagement.
• It operationalizes Peer Learning with Close Peer and Distant Peer profiles to co-construct playlists, showing significant improvements in interest expansion and value.
• Experimental results indicate that Close Peer conditions yield higher engagement and lower variability compared to baseline, while Distant Peer elicits more diverse responses.

Peer Recommendation: Fostering Engagement in Collaborative Dialogue-Based Recommender Systems

Introduction

The paper "Beyond Serendipity: From Exposing the Unknown to Fostering Engagement through Peer Recommendation" (2604.16818) positions recommendation not as a unidirectional, user-passive process but as a collaborative, dialogue-driven engagement. By operationalizing the theory of Peer Learning, the authors design a system where the user and an LLM-powered Peer agent mutually act as recommenders and recipients, co-constructing a playlist in natural language dialogue. This framework aims to move beyond static notions of serendipity—mere exposure to unfamiliar items—toward fostering user engagement, interest expansion, and meaningful interaction through explicit "otherness" in agent persona.

Conceptual Framework and System Architecture

The Peer Recommendation system leverages concepts from Peer Learning to transform recommender models into active, conversational agents with distinct, and critically, non-user-aligned preferences. The central tenet is that mutual recommendation in dialogue, rather than simple algorithmic exposure, is instrumental for genuine cognitive engagement with unfamiliar content.

Figure 1: Conceptual diagram illustrating the Peer Recommendation scenario, where user and Peer collaboratively recommend items via natural-language dialogue.

The architecture is composed of four stages:

1. User Profile Generation: LLMs synthesize extended user profiles from demographics and playlist data, describing musical values and behavioral tendencies.
2. Peer Profile Generation: From this user profile, the system algorithmically constructs a Peer agent, which can be either (a) a Close Peer, sharing core preferences but encouraging adjacent exploration, or (b) a Distant Peer, establishing significant taste divergence while retaining minimal commonality.
3. Mutual Recommendation via Dialogue: Iterative, chat-based exchanges are mediated by the LLM, with each message processed through a three-step flow—intent and reception analysis, action candidate selection, and persona-grounded response synthesis.

   Figure 2: Overview of the end-to-end Peer Recommendation pipeline, including user and Peer profile creation and the collaborative playlist construction cycle.

   

   Figure 3: Conversation management workflow demonstrating the multi-stage processing of each user input via intent analysis, candidate selection, and persona-grounded response generation.

4. Collaborative Playlist Construction: Only items mutually endorsed by user and Peer are included in the synthesized playlist, ensuring bilateral engagement.

Crucially, both user and Peer profiles are static within a session, removing confounds from dynamic adaptation and focusing analytic leverage on the Peer preference distance variable.

Experimental Design and Quantitative Evaluation

A within-subjects design (N=14) contrasts three system configurations—Close Peer, Distant Peer, and Baseline (lacking a distinct persona). Each participant recommends songs to their Peer agent, receiving recommendations in return and curating a joint playlist.

Key dependent variables:

• Perceived interest expansion (custom single-item measure)
• Intrinsic Motivation Inventory (IMI) subscales: Value and Tension/Pressure

The results indicate statistically significant improvements over Baseline in both interest expansion and perceived value when users interact with Close Peer conditions. Specifically, Close Peer achieves higher means and lower variance than Baseline for both metrics—$M = 5.21$ vs. $M = 4.43$ for interest expansion ($p = .044$, $r = .46$); $M = 5.62$ vs. $M = 4.55$ for IMI Value ($p = .006$, $r = .67$).

Figure 4: Comparative distribution of perceived interest expansion and IMI Value across experimental conditions; Close Peer yields higher means and reduced variance.

The Distant Peer condition does not show aggregate significance but presents higher inter-individual variability, as reflected in open comments. Notably, the Baseline agent is consistently criticized as being excessively accommodating, lacking the productive friction that induces cognitive engagement.

Qualitative Analysis and Individual Differences

Thematic analysis identifies two participant archetypes:

• Exploration-Oriented Users: Prefer large preference distances and derive greater value from Distant Peer interactions. They cite the necessity of agent disagreement and "pushback" for stimulating discovery, viewing accommodation as a deficit.
• Affinity-Oriented Users: Value alignment and a sense of being understood, finding Distant Peer burdensome. For these users, Close Peer or even Baseline is preferable.

These data demonstrate that perceived benefit from agent "otherness" is non-uniform and contingent on individual exploratory readiness. Absence of significant correlation between subjective Peer similarity and perceived interest expansion further decouples alignment from engagement.

Implications for Recommender Systems Design

The findings substantiate that explicit agent persona design—especially constructed "otherness"—is a key driver of engagement and self-reported value in collaborative recommendation. Unlike traditional LLM-based assistants that mirror user preferences, agents with well-calibrated preference divergence can ignite exploration and productive dialogue. Theoretical implications extend to human-centered AI, advocating for adaptive, persona-driven recommenders that modulate preference distance in response to user traits and state.

Practically, this suggests interactive systems where users or agents dynamically tune their mutual "distance," possibly through meta-dialogic negotiation, to optimize challenge and familiarity. The approach generalizes beyond music to any collaborative filtering domain involving list curation and social exploration, e.g., news, educational resources, and literary recommendation.

Conclusion

This work offers compelling empirical and conceptual evidence for positioning dialogue-based, persona-driven Peer agents at the center of next-generation recommender systems. Agents endowed with explicit and differentiated profiles, particularly those diverging from the user, significantly enhance engagement and interest expansion for exploration-oriented users. Contradictory preference distances, however, require adaptive mediation as they may be counterproductive for affinity-oriented individuals. These results highlight the need for user modeling that incorporates not only static taste profiles but also attitudes toward cognitive challenge and engagement.

Future developments may explore adaptive control mechanisms for preference distance, real-time adjustment of agent behavior, and richer socio-cognitive modeling in dialogue systems to better align recommender strategies with user disposition and engagement goals.