ATH-MaaS/Marco-DeepResearch

What it solves

Marco DeepResearch is a comprehensive initiative designed to bridge the gap between AI agents and human experts in complex, real-world tasks. It addresses critical limitations in domain-specific reasoning, hierarchical rule application, and large-scale information seeking, particularly in vertical domains like trade, legal, and medical fields.

How it works

The project consists of a suite of benchmarks, frameworks, and models that target different agent capabilities:

• Table-as-Search: Formulates long-horizon search as a table completion task, using a hierarchical multi-agent framework to coordinate "Wide" (tabular) and "Deep" (multi-hop) search agents.
• UMEM (Unified Memory Extraction and Management): A self-evolving memory system that optimizes how agents extract and manage insights from interaction traces to avoid rote memorization.
• Search2Skill: A framework that uses rubric-based reinforcement learning to distill reusable expert-domain skills from external search results.
• Benchmarks (HSCodeComp & DeepWideSearch): Specialized datasets used to evaluate an agent's ability to apply hierarchical rules (e.g., for tariff classification) and perform deep-and-wide information gathering.

Who it’s for

This project is for AI researchers and developers building production-ready agents for complex business intelligence, cross-border trade compliance, and automated customer service auditing.

Highlights

• Production-Ready Frameworks: Includes Table-as-Search for high-recall business development intelligence.
• Self-Evolving Memory: UMEM allows agents to learn from expert ground truth and update their behavior in minutes rather than days.
• Specialized Benchmarks: HSCodeComp provides a rigorous test for hierarchical rule application in vertical domains.
• Efficient Model: Includes the Marco-DeepResearch-8B model specifically tuned for deep research agent tasks.