The Agent Perspective In LLM-Based Strategic Information Retrieval Ecosystems

Abstract: Information retrieval ecosystems are strategic environments in which publishers, mediators, advertisers, and users interact under mechanisms that allocate visibility and revenue. The integration of large language models (LLMs) into search and question answering systems fundamentally reshapes these dynamics, altering ranking incentives and long-term corpus evolution. My dissertation develops a unified strategic perspective on LLM-based information retrieval, focusing on the publisher’s viewpoint under undisclosed mediation mechanisms. We first study competitive search under diversity-based ranking grounded in LLM-derived semantic representations. While standard relevance-based ranking induces herding behavior --- where publishers mimic top-ranked documents --- semantic diversification penalizes excessive similarity. We provide the first theoretical and empirical analysis of this setting, proving the existence of a min-max regret equilibrium and showing that diversification changes strategic incentives. In equilibrium, some publishers target alternative rank positions, mitigating herding and improving corpus diversity. To enable scalable experimentation, we also introduce LEMSS, an LLM-based multi-agent competitive search simulator. LEMSS models LLMs as strategic publishers interacting under configurable ranking functions. Experiments reveal that LLM-based agents reproduce and often amplify convergence behaviors observed in human competitions, demonstrating that generative models act as active strategic participants. We then propose Reinforcement Learning from Ranker Feedback (RLRF), a framework for training LLMs to become ranker-aware competitive agents. Using synthetic competition data, we fine-tune LLMs through Direct Policy Optimization, enabling them to internalize ranking signals and adapt strategically. RLRF-trained agents outperform prompt-based baselines and formalize two strategic learning modes: learning from signals from the ranking function and adapting to successful competitors. Finally, we extend the framework to Sponsored Question Answering, designing an auction-based mechanism for allocating sponsored content within generated responses. We establish equilibrium and welfare properties, bridging classical sponsored search theory with LLM-mediated QA systems. Together, these contributions provide a game-theoretic and algorithmic foundation for designing strategic, diversity-aware, and welfare-aligned retrieval systems in the era of generative AI.