Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Safe and Human-Aligned Game Conversational Recommendation via Multi-Agent Decomposition

Zheng Hui, Xiaokai Wei, Yexi Jiang, Kevin Gao, Chen Wang, Frank Ong, Se-eun Yoon, Rachit Pareek, Michelle Gong

TL;DR

This paper introduces MATCHA, a modular multi-agent framework for safe and human-aligned game conversational recommendations. It decomposes the CRS pipeline into specialized agents for intent parsing, tool-augmented candidate generation, multi-LLM ranking with reflection, risk control, and explainability to address complex game constraints, rapid content drift, and safety risks. Across real-user Roblox data and cross-domain assessments, MATCHA achieves superior relevance, diversity, and factuality while maintaining strong adversarial robustness and transparent explanations; ablations confirm the necessity of each module. The work demonstrates practical deployment feasibility in an internal setting and provides actionable guidance for extending multi-agent, safety-conscious CRS to diverse domains.

Abstract

Conversational recommender systems (CRS) have advanced with large language models, showing strong results in domains like movies. These domains typically involve fixed content and passive consumption, where user preferences can be matched by genre or theme. In contrast, games present distinct challenges: fast-evolving catalogs, interaction-driven preferences (e.g., skill level, mechanics, hardware), and increased risk of unsafe responses in open-ended conversation. We propose MATCHA, a multi-agent framework for CRS that assigns specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking with reflection, explanation, and risk control which enabling finer personalization, long-tail coverage, and stronger safety. Evaluated on real user request dataset, MATCHA outperforms six baselines across eight metrics, improving Hit@5 by 20%, reducing popularity bias by 24%, and achieving 97.9% adversarial defense. Human and virtual-judge evaluations confirm improved explanation quality and user alignment.

Toward Safe and Human-Aligned Game Conversational Recommendation via Multi-Agent Decomposition

TL;DR

This paper introduces MATCHA, a modular multi-agent framework for safe and human-aligned game conversational recommendations. It decomposes the CRS pipeline into specialized agents for intent parsing, tool-augmented candidate generation, multi-LLM ranking with reflection, risk control, and explainability to address complex game constraints, rapid content drift, and safety risks. Across real-user Roblox data and cross-domain assessments, MATCHA achieves superior relevance, diversity, and factuality while maintaining strong adversarial robustness and transparent explanations; ablations confirm the necessity of each module. The work demonstrates practical deployment feasibility in an internal setting and provides actionable guidance for extending multi-agent, safety-conscious CRS to diverse domains.

Abstract

Conversational recommender systems (CRS) have advanced with large language models, showing strong results in domains like movies. These domains typically involve fixed content and passive consumption, where user preferences can be matched by genre or theme. In contrast, games present distinct challenges: fast-evolving catalogs, interaction-driven preferences (e.g., skill level, mechanics, hardware), and increased risk of unsafe responses in open-ended conversation. We propose MATCHA, a multi-agent framework for CRS that assigns specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking with reflection, explanation, and risk control which enabling finer personalization, long-tail coverage, and stronger safety. Evaluated on real user request dataset, MATCHA outperforms six baselines across eight metrics, improving Hit@5 by 20%, reducing popularity bias by 24%, and achieving 97.9% adversarial defense. Human and virtual-judge evaluations confirm improved explanation quality and user alignment.
Paper Structure (46 sections, 3 equations, 6 figures, 8 tables)

This paper contains 46 sections, 3 equations, 6 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Conversational Recommendation System
  4. Risk factor in LLM-based Dialogue System
  5. Problem Formulation
  6. MATCHA Framework
  7. Risk Control
  8. Jailbreak Prevention Agent
  9. Dangerous Content Detection
  10. Candidate Generation
  11. Ranking and Reflection
  12. Ranking Agent
  13. Reflection Agent
  14. Explainability
  15. Experiments Setups
  16. ...and 31 more sections

Figures (6)

  • Figure 1: Overview of Our MATCHA framework. The system processes user requests through safety agents, generates game candidates using diverse tools, refines them via ranking and reflection agents, and provides final recommendations with detailed explanations.
  • Figure 2: Ablation study results for the Jailbreak Prevention Agent on the WildJailbreak dataset.
  • Figure 3: Ablation study results demonstrating the impact of removing the Reflection, multi-LLM collaboration for decision-making, and tool-used.
  • Figure 4: Demonstration of the application interface showcasing recommendations with reasons.
  • Figure 5: Illustration of the evaluation interface used for human annotations, showcasing the layout and components visible to annotators.
  • ...and 1 more figures