LLM$\times$MapReduce-V3: Enabling Interactive In-Depth Survey Generation through a MCP-Driven Hierarchically Modular Agent System

arXiv 2025 47.8 method

TLDR

A hierarchically modular MCP-based agent system for interactive in-depth survey generation, outperforming baselines in human evaluations.

Reasoning

Strengths include a modular MCP architecture enabling human-in-the-loop control and dynamic workflow orchestration. Weaknesses are the narrow focus on survey generation without addressing broader scientific discovery or evaluating factual accuracy.

Read-first score

Read-first score 47.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation

Topical relevance 42%

43.3

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 55.

Keyword Scores

survey generation

literature review agent

paper writing agent

automated research

research automation

AI for scientific research

autonomous research agent

AI scientist

scientific discovery agent

automated scientific discovery

automated experimentation

experiment design agent

Deep Analysis

Innovations

MCP-driven hierarchically modular agent system for long-form survey generation
Decomposition of functional components (skeleton initialization, digest construction, skeleton refinement) as independent MCP servers that can be aggregated into higher-level servers
High-level planner agent that dynamically orchestrates workflow by selecting modules based on MCP tool descriptions and execution history
Human-in-the-loop intervention enabled by modular decomposition, allowing user control and customization through multi-turn interaction

Methodology

The system employs a multi-agent architecture where functional components are implemented as independent MCP servers, which can be aggregated hierarchically. A high-level planner agent dynamically selects modules using their MCP tool descriptions and execution history, while multi-turn human interaction captures research perspectives to generate a comprehensive skeleton that is then developed into a full survey. Human evaluations compare the system against representative baselines on content depth and length.

Key Results

Human evaluations demonstrate that the system surpasses representative baselines in both content depth and length.

Resources

Paper