Spark Framework Enables Task-Specific Search Personalization with Coordinated Large Language Model Agents

Personalized search continually strives to understand the complex and changing information needs of individual users, a task that proves difficult for systems relying on fixed user profiles or inflexible processes. To address this challenge, Gaurab Chhetri, Subasish Das, and Tausif Islam Chowdhury, all from Texas State University, present SPARK, a novel framework that uses coordinated, persona-based language model agents to deliver highly specific and adaptable search results. This innovative approach formalises a detailed ‘persona space’ encompassing roles, expertise, and task context, and introduces a ‘Persona Coordinator’ which intelligently activates the most appropriate agent for each query. By drawing on principles of cognitive architecture and multi-agent systems, SPARK not only models how personalised search emerges from collaborative agent behaviour, but also offers a pathway towards future search systems that more accurately reflect the nuanced and dynamic nature of human information seeking.

Agent-Driven Search Personalization Balances Privacy and Accuracy

Scientists developed SPARK, a system that uses multiple agents to improve search personalization, addressing limitations in traditional methods prone to privacy leaks and inaccuracies. This multi-agent system retrieves information, reasons about user needs, shares knowledge, and adapts to both long-term and short-term preferences, offering a more robust and nuanced approach to understanding user intent. SPARK leverages Retrieval-Augmented Generation (RAG) to ground responses in external knowledge, reducing the risk of generating false information, and employs techniques like Plan-and-Solve to guide complex reasoning processes. The system prioritizes privacy through data minimization and privacy-preserving techniques, ensuring user information is protected during operation. Researchers introduced SafeAgentBench, a benchmark for evaluating the safety, robustness, and coordination of multi-agent large language model (LLM) systems, and drew upon related areas such as personalized search, user modeling, and recommender systems to build this innovative framework. SPARK’s design emphasizes adaptability, scalability, and safety, offering a significant advancement in personalized search technology.

Agent-Driven Personalization with Dynamic Persona Spaces

The research team engineered SPARK, a novel framework for personalized search that moves beyond static user profiles, employing coordinated, persona-based large language model (LLM) agents to deliver task-specific personalization. Central to SPARK is a formalization of a ‘persona space’, defined by role, expertise, task context, and domain, enabling highly specialized agent behavior and context-aware retrieval. A Persona Coordinator dynamically interprets queries and activates the most relevant agents from a pool of possibilities, ensuring nuanced analysis and efficient information gathering. Each agent operates independently, supported by dedicated long- and short-term memory stores, allowing it to retain information across sessions and within the current task. Structured communication protocols, including shared memory, iterative debate, and relay-style knowledge transfer, facilitate collaboration between agents, mirroring cognitive architectures and multi-agent coordination theory. This approach captures the complexity and fluidity of human information-seeking behavior, offering a significant advancement over conventional search architectures.

Dynamic Search via Multi-Agent Personas

Scientists developed SPARK, a novel framework for personalized search that achieves a dynamic and adaptive approach to information retrieval by moving beyond static user profiles. The work introduces a formal persona model, defining agents by role, expertise, task context, and domain, enabling fine-grained specialization of search behavior. A central Persona Coordinator analyzes incoming queries and activates the most relevant agents, facilitating a system where multiple specialized agents collaborate to retrieve and synthesize information. SPARK’s multi-agent architecture supports diverse coordination protocols, including independent operation with fused results, relay-style knowledge transfer, and constrained debate overseen by a judge agent. Each agent maintains both short-term working memory and long-term semantic memory, mirroring cognitive architectures and allowing the system to distinguish between transient context and persistent user interests. Researchers propose testable hypotheses regarding coordination efficiency and personalization quality, evaluating these predictions using both offline simulations and user studies.

SPARK, Persona-Based Search and Adaptive Refinement

This research presents SPARK, a novel framework that reimagines search personalization as a coordinated effort between specialized agents, each embodying a distinct persona. The team formalizes the concept of a ‘persona space’ encompassing role, expertise, task context, and domain, and demonstrates how a ‘Persona Coordinator’ can dynamically activate the most appropriate agent for a given query. By integrating principles from cognitive architectures and multi-agent systems, SPARK achieves context-sensitive retrieval and synthesis, while also supporting continuous refinement of these personas through adaptive learning mechanisms. The framework delivers both relevant and diverse search results by combining stochastic routing, calibrated fusion, and objectives focused on understanding user intent. While the architecture introduces some coordination overhead, adaptive gating and protocol selection effectively mitigate this cost, enabling efficient personalization. Future work will explore live-system deployments, advanced fusion models, and strategies for correcting long-term memory inaccuracies, ultimately aiming to build more robust and adaptable search experiences.