Technical Summary: MACRDR – Enhancing Interest Discovery and Diversity in News Recommendation via Multi-Agent Reflection

1. Problem Statement

Personalized news recommendation systems frequently suffer from "filter bubbles" and "echo chambers" due to an excessive reliance on historical user interaction patterns. While deep learning-based approaches have improved semantic modeling and click prediction accuracy, they often reinforce existing preferences, rendering them insensitive to dynamic shifts in user intent and emerging interests.

Existing diversity-aware methods typically address this through heuristic reranking (e.g., Maximum Marginal Relevance) or representation optimization (e.g., contrastive learning). However, these approaches primarily focus on item-level distribution adjustments or feature separation without explicitly modeling the underlying evolution process of user intent. Consequently, they may introduce diverse content that does not align with a user's actual latent or emerging interests. Furthermore, while Large Language Models (LLMs) offer strong reasoning capabilities, applying a single LLM agent to complex recommendation tasks often leads to ranking instability, popularity bias, and hallucinations due to limited long-term context understanding and the lack of persistent memory mechanisms.

2. Methodology: MACRDR Framework

The authors propose MACRDR (Multi-Agent Collaboration and Reflection for Diversity-aware Recommendation), a framework that reformulates news recommendation as a dynamic intent calibration process rather than a static ranking optimization problem. The framework decouples the process into two stages: Base Retrieval and Cognitive Reranking, mediated by a multi-agent system with a predictive-reflection feedback loop.

2.1 Core Architecture

The framework consists of three primary modules:

1. Base Recommendation Module: Acts as the candidate generation stage, utilizing a deep learning model (specifically a PLM-Attention network in the experiments) to encode user history and news content. It generates a high-relevance candidate set ($C_{cand}$) with initial ranking scores, serving as a constraint for subsequent agent-based reranking to ensure baseline accuracy.
2. Dynamic Profile Memory Module: Designed to overcome the lack of persistent memory in LLM agents, this module stores structured, version-controlled user profiles ($Prof_u^v$). Unlike implicit vector representations, it uses lightweight, document-based (JSON) structured memory containing:
   • Basic Attributes: Core interest categories and persona descriptions.
   • Cognitive Adjustment Patches ($Patch_u^v$): Explicit natural language instructions generated by reflection agents to guide future adjustments.
   • Evolution Metadata: Records of profile versions, update triggers, and timestamps to maintain temporal consistency.
3. Multi-Agent Collaboration Module: The decision-making core comprising four specialized agents:
   • Manager Agent: Orchestrates task routing, state management, and adaptive distribution of tasks to other agents based on system state.
   • UserAnalyst Agent: Initializes user profiles by extracting interpretable preference information from historical interactions using LLM summarization.
   • Action Agent: Performs two roles: (1) predicting potential user interests based on the current profile during the offline evolution stage, and (2) performing final intent-aware reranking during the online stage by balancing base relevance scores with reflection guidance.
   • Reflect Agent: Compares the Action Agent's predicted interests ($\hat{Y}_t$) with actual user feedback ($Y_t$). It calculates semantic discrepancy ($\delta_t$) and generates explicit cognitive adjustment patches if the deviation exceeds a dynamic tolerance threshold ($\theta$).

2.2 The Predictive-Reflective Feedback Loop

The core innovation lies in the offline profile evolution process:

1. Prediction: The Action Agent predicts expected news subcategories based on the current user profile.
2. Reflection: The Reflect Agent compares predictions against actual user interactions.
3. Calibration: If the deviation is significant ($\delta_t > \theta$), the Reflect Agent analyzes the cause and generates a "cognitive patch" ($Patch_u^{v+1}$). This patch serves as an explicit instruction for future recommendations, guiding the system to explore emerging interests while correcting for prediction errors.
4. Online Reranking: When a recommendation request arrives, the Action Agent retrieves the latest profile and candidate set. It optimizes a joint objective: maximizing relevance (from the base model) while incorporating the exploration instructions from the cognitive patches.

3. Key Contributions

The paper outlines three primary contributions:

1. Reformulation of Recommendation: MACRDR shifts the paradigm from static ranking to dynamic intent calibration. It decouples base retrieval (accuracy-focused) from cognitive reranking (diversity-focused), leveraging the reasoning capabilities of LLM-based agents while retaining the efficiency of traditional retrieval models.
2. Predictive-Reflective Collaboration Mechanism: The authors design a mechanism that explicitly models user intent evolution. By comparing predicted interests with real feedback, the system generates "cognitive patches" that guide recommendation adjustments, allowing for the capture of emerging intents rather than just historical patterns.
3. Dynamic Profile Memory: A version-controlled memory module is introduced to preserve the trajectory of user preference changes and reflective corrections. This enables long-term intent tracking and provides interpretable, structured context for multi-agent reasoning, addressing the memory limitations of standard LLMs.

4. Experimental Results

The framework was evaluated on the MIND (Microsoft News Dataset) using a subset of 2,000 active users (MIND-2000) with a temporal split strategy to simulate natural interest evolution.

4.1 Performance Metrics

The evaluation considered both accuracy (AUC, MRR, NDCG@K) and diversity/interest discovery (Interest Discovery Rate - IDR, Serendipity - SR, Intra-List Diversity - ILD, Interest Coverage - IC, Aggregate Coverage - AC).

4.2 Key Findings

• Accuracy vs. Diversity Trade-off: MACRDR maintained competitive recommendation accuracy (e.g., NDCG@10 of 0.4087) comparable to advanced LLM-based baselines like PNR-LLM and LKPNR, despite introducing exploratory items.
• Superior Interest Discovery: MACRDR achieved the highest Interest Discovery Rate (IDR@5: 26.43%) and Serendipity (SR@5: 0.0853) among all baselines, significantly outperforming deep learning models (e.g., NRMS, NAML) and other LLM-based approaches.
• Ablation Studies:
  • Removing the Base Recommender caused a drastic drop in accuracy, confirming the necessity of a high-quality relevance-constrained candidate set.
  • Removing the Reflect Agent or Action Agent reduced the system's ability to dynamically calibrate preferences, leading to lower IDR and SR scores.
  • Removing Rank Constraints (relying solely on LLM reranking) improved diversity metrics but significantly degraded accuracy, validating the need to preserve base relevance signals.
• Parameter Sensitivity: The study found that a historical interaction slice size of 3 and a dynamic tolerance threshold ($\theta$) of 6 provided the optimal balance between stability and exploration.

5. Significance and Claims

The paper claims that MACRDR provides an effective and interpretable solution for mitigating filter bubbles in news recommendation. Its significance lies in:

• Modeling Intent Evolution: Unlike methods that treat diversity as a post-hoc constraint, MACRDR models the process of intent evolution through a reflection loop, allowing the system to distinguish between stable preferences and emerging interests.
• Controlled Exploration: By using "cognitive patches" derived from prediction errors, the system performs targeted exploration rather than random diversification, ensuring that recommended diverse content is actually relevant to the user's latent needs.
• Practical Feasibility: The architecture demonstrates that multi-agent collaboration with structured memory can effectively handle complex recommendation tasks, balancing the reasoning power of LLMs with the computational efficiency and accuracy of traditional deep learning models.

The authors conclude that while the framework shows promise, future work could involve using simulated user agents to incrementally train the base model and integrating multimodal information for finer-grained intent modeling.