Breaking User-Centric Agency: A Tri-Party Framework for Agent-Based Recommendation

📝 Paper Summary

Agent-based Recommender Systems Multi-Stakeholder Fairness

TriRec transforms items from passive entities into active agents that generate personalized self-promotion, regulated by a platform agent to balance user relevance, item survival, and fairness.

Core Problem

Current agent-based recommenders are predominantly user-centric, optimizing user utility while treating items as passive, which leads to exposure concentration (Matthew effect) and reduced creator incentives.

Why it matters:

The 'Matthew effect' concentrates exposure on popular items, leaving long-tail creators invisible and causing creator churn due to lack of incentives
Excessive optimization for short-term user engagement homogenizes the content pool, eroding supply-side diversity and harming the platform's long-term ecosystem health
Prior multi-agent methods focus on user utility, treating item agents merely as information carriers rather than active stakeholders advocating for their own visibility

Concrete Example: A CD player might be generically recommended based on category, but under this system, it actively emphasizes 'high audio fidelity' to a musician user versus 'easy playback' to a senior user to win exposure.

Key Novelty

Tri-party LLM-agent Recommendation (TriRec)

**Item Agency:** Unlike passive item embeddings, item agents actively generate personalized 'self-promotion' content tailored to specific users to advocate for their own exposure
**Tri-Party Alignment:** Explicitly coordinates three distinct utility functions: User (relevance), Item (exposure/survival), and Platform (fairness/stability)
**Decoupled Regulation:** A two-stage architecture where Stage 1 maximizes semantic matching via promotion, and Stage 2 regulates the global list for fairness constraints

Architecture

The TriRec framework overview, illustrating the two-stage pipeline: (1) Generative item self-promotion and (2) Platform-led multi-objective re-ranking.

Evaluation Highlights

Item self-promotion simultaneously increased average exposure and click-through probability, challenging the trade-off assumption between fairness and effectiveness
The introduction of personalized persuasive content enhanced recommendation accuracy on the user side compared to passive item descriptions

Breakthrough Assessment

8/10

Significant conceptual shift from user-centric to multi-stakeholder agency in LLM-based recommendation. The idea of items actively 'campaigning' for exposure via generated text is novel.

⚙️ Technical Details

Problem Definition

Setting: Multi-stakeholder agent-based recommendation involving Users, Items, and Platform

Inputs: User interaction history/memory, Item metadata, Platform fairness constraints

Outputs: Ranked list of items optimized for tri-party utility (Relevance + Exposure + Fairness)

Pipeline Flow

Item Agent (Generates personalized self-promotion)
User Agent (Evaluates relevance based on promotion + memory)
Platform Agent (Re-ranks candidate list for fairness/utility)

System Modules

Item Agent (Stage 1: Generative Self-Promotion)

Generate user-specific promotional text to overcome cold-start and improve matching

Model or implementation: Unspecified frozen pre-trained LLM

User Agent (Stage 1: Generative Self-Promotion)

Evaluate candidates based on generated promotion and internal preference memory

Model or implementation: Unspecified frozen pre-trained LLM

Platform Agent

Sequential re-ranking of the candidate list to balance relevance, item utility, and fairness

Model or implementation: Unspecified frozen pre-trained LLM

Novel Architectural Elements

Introduction of 'Item Agency' where items actively generate dynamic content (self-promotion) during inference rather than having static representations
Tri-party utility alignment via a sequential agentic re-ranking stage (Stage 2) that acts as a platform regulator

Modeling

Base Model: Unspecified frozen pre-trained LLM (Specific architecture like Llama/GPT not named in provided text)

Training Method: AgentCF-style interaction training (offline replay of historical interactions)

Objective Functions:

Purpose: Maximize User Utility (Relevance).

Formally: hat{r}(u,i) based on agent semantic similarity.
Purpose: Maximize Item Utility (Exposure).

Formally: Expected Item Utility (EIU) aggregating position-dependent exposure v(k) and predicted CTR.
Purpose: Maximize Platform Utility (Fairness).

Formally: Minimize Distributional Group Unfairness (DGU) and Maximal Group Unfairness (MGU).

Compute: Not reported in the paper

Comparison to Prior Work

vs. AgentCF: TriRec adds Item Agency (active promotion) and Platform Agency (fairness regulation), whereas AgentCF focuses on User-Item relevance simulation
vs. DualRec: DualRec treats items as queries for optimization; TriRec empowers them with generative agency to create persuasive content
vs. HLLM-Creator: HLLM-Creator uses LLMs for content augmentation; TriRec integrates this into a tri-party negotiation framework where items actively compete

Limitations

Relies on the reasoning capability of the underlying frozen LLM
Sequential re-ranking may introduce latency compared to non-agentic scoring
Specific quantitative benchmarks and baselines are not detailed in the provided text snippet

Reproducibility

Code: https://github.com/Marfekey/TriRec

Code is publicly available at https://github.com/Marfekey/TriRec. The specific LLM backbone used for agents is not named in the provided text snippet. Datasets are described generically as 'public datasets'.

📊 Experiments & Results

Evaluation Setup

Re-ranking and Recommendation simulation

Benchmarks:

Public datasets (unspecified in text) (Recommendation / Fairness evaluation)

Metrics:

Accuracy (User Utility)
Distributional Group Unfairness (DGU)
Maximal Group Unfairness (MGU)
Expected Item Utility (EIU)
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Incorporating item self-promotion leads to an increase in both average exposure and click-through probability, challenging the conventional view that fairness hurts accuracy
Personalized persuasive content generated by item agents improves matching quality (accuracy) compared to static item descriptions
The platform agent successfully regulates the trade-off between user relevance and long-term ecosystem health (fairness/item survival)

📚 Prerequisite Knowledge

Prerequisites

Recommender Systems (Collaborative Filtering, Ranking)
LLM-based Agents
Multi-objective Optimization

Key Terms

Matthew effect: The phenomenon where popular items get more popular while less popular ones (long-tail) lose visibility, leading to rich-get-richer dynamics

Cold-start: The difficulty of recommending new items or items with few historical interactions due to lack of data

Agentic Recommendation: Using LLM-based autonomous agents to simulate user and item behaviors for preference modeling and interaction

Self-promotion: The mechanism where an item agent actively generates text to persuade a specific user agent, rather than just being a passive candidate

DGU: Distributional Group Unfairness—a metric measuring the discrepancy between item group exposure in recommendations versus their historical distribution

MGU: Maximal Group Unfairness—the maximum unfairness observed across any single group

EIU: Expected Item Utility—a metric quantifying the realized gain for an item based on its position in the ranking and predicted click-through rate

AgentCF: A foundational agent-based collaborative filtering framework (cited as prior work) where agents simulate interactions to refine preferences