MAPS: Modular Agentic Planning and Search

Kayla Lee

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-69
May 15, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-69.pdf

We introduce MAPS, a modular system for multi-hop retrieval and question answering in open-domain settings. In open-book, source-not-provided scenarios, systems must iteratively retrieve and reason over documents without knowing in advance which sources contain the answer. MAPS addresses this challenge through a structured workflow composed of three dedicated components: a Query Planner that generates and refines search strategies, an Article Selector that filters relevant results, and a Content Extractor that identifies concise supporting spans. Each module is implemented as a lightweight, task-specific agent, enabling modular separation that enhances robustness, scalability, and interpretability. Specifically, MAPS (1) avoids prompt overflow by distributing tasks, (2) enables concurrent document processing to overcome sequential bottlenecks, (3) supports dynamic query refinement through inter-module feedback, and (4) achieves strong performance using smaller models without incurring additional compute cost. We evaluate both the baseline and our approach on a 1,000-question subset of the HotpotQA dataset and achieve a 72% success rate with a GPT-based judge — substantially outperforming the baseline at 42.1%.

\"Edit"; ?>

BibTeX citation:

@mastersthesis{Lee:EECS-2025-69,
    Author = {Lee, Kayla},
    Title = {MAPS: Modular Agentic Planning and Search},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-69.html},
    Number = {UCB/EECS-2025-69},
    Abstract = {We introduce MAPS, a modular system for multi-hop retrieval and question answering in open-domain settings. In open-book, source-not-provided scenarios, systems must iteratively retrieve and reason over documents without knowing in advance which sources contain the answer. MAPS addresses this challenge through a structured workflow composed of three dedicated components: a Query Planner that generates and refines search strategies, an Article Selector that filters relevant results, and a Content Extractor that identifies concise supporting spans. Each module is implemented as a lightweight, task-specific agent, enabling modular separation that enhances robustness, scalability, and interpretability. Specifically, MAPS (1) avoids prompt overflow by distributing tasks, (2) enables concurrent document processing to overcome sequential bottlenecks, (3) supports dynamic query refinement through inter-module feedback, and (4) achieves strong performance using smaller models without incurring additional compute cost. We evaluate both the baseline and our approach on a 1,000-question subset of the HotpotQA dataset and achieve a 72% success rate with a GPT-based judge — substantially outperforming the baseline at 42.1%.}
}

EndNote citation:

%0 Thesis
%A Lee, Kayla
%T MAPS: Modular Agentic Planning and Search
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 15
%@ UCB/EECS-2025-69
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-69.html
%F Lee:EECS-2025-69