ATH-MaaS/Marco-DeepResearch

What it solves

Marco DeepResearch addresses the gap between AI agents and human experts in complex, real-world tasks. It specifically targets limitations in domain-specific reasoning (such as legal or medical), hierarchical rule application, large-scale information seeking (both deep and wide), and the ability of agents to evolve their skills and memory without overfitting.

How it works

The project provides a suite of benchmarks and frameworks to improve agent capabilities:

• Table-as-Search: A multi-agent framework that treats long-horizon search as a table completion task, using hierarchical orchestration of "Wide" and "Deep" sub-agents to track state and plan effectively.
• 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: Includes HSCodeComp for testing hierarchical rule application in trade tariffs and DeepWideSearch for evaluating deep-and-wide information seeking.

Who it’s for

This is designed for AI researchers and developers building production-ready agents for vertical domains (e.g., e-commerce, trade, legal) and those working on long-horizon planning and self-evolving agent memory.

Highlights

• Self-Evolving Capabilities: UMEM reduces tuning time for customer service agents from days to minutes by autonomously learning from expert gaps.
• High-Performance Agent: Includes the Marco-DeepResearch-8B model and a specialized agent orchestrator.
• Domain-Specific Success: Demonstrates significant gains in HS code classification for cross-border trade compared to general-purpose LLMs.
• Structured Planning: Uses the "Table-as-Search" approach to improve entity recall and attribute completeness in business intelligence tasks.