ANAGENT For Enhancing Scientific Table & Figure Analysis

       

---

📊 ANABENCH: Benchmarking Scientific Table & Figure Analysis

ANABENCH is a large-scale benchmark featuring 63,178 instances from 9 scientific domains and 170 fine-grained subdomains, systematically categorized along 7 complexity dimensions to evaluate scientific table and figure analysis capabilities.

Key Features

• 📈 63,178 Instances across diverse scientific domains
• 🔬 9 Broad Domains: Computer Science, Electrical Engineering, Mathematics, Physics, Economics, Quantitative Biology, Quantitative Finance, Statistics, Biomedicine
• 🎯 170 Fine-grained Subdomains for more comprehensive coverage
• 📐 7 Complexity Dimensions: Type, Domain, Format, Source, Width, Depth, Objective

Complexity Dimensions

1. Type: Tables, figures, or combined multimodal contents
2. Domain: 9 broad scientific fields with 170 fine-grained subdomains
3. Format: LaTeX, XML, and other representations
4. Source: General research papers vs. review & survey papers
5. Width: Self-contained, internal, external, or mixed reference scope
6. Depth: Shallow description vs. in-depth inferential analysis
7. Objective: Methodology-oriented vs. experiment-oriented analysis

---

🌟 ANAGENT: Multi-Agent Collaborative Scientific Analysis

ANAGENT is a multi-agent framework that simulates human research workflows through four specialized agents working collaboratively to perform high-quality scientific table and figure analysis.

Specialized Agents

• 🎯 PLANNER: Decomposes complex tasks into actionable subtasks and provide task-oriented planning guidance
• 🔍 EXPERT: Performs iterative knowledge acquisition through specialized tool execution and targeted retrieval
• 🖋️ SOLVER: Synthesizes accumulated knowledge to generate coherent, context-aware analysis solutions
• 🔧 CRITIC: Assesses analysis quality through five-dimensional evaluation and provides targeted refinement feedback

---

🛠️ 1. Experiment Setup

Installation

# Coming soon

---

🚀 2. Quick Start

# Coming soon

---

📈 3. Modular Optimization

Supervised Finetuning

# Coming soon

---

📝 Citation

Thanks much for your interests!

If you find this work useful, please kindly cite:

@article{guo2026anagent,
  title={ANAGENT For Enhancing Scientific Table & Figure Analysis},
  author={Guo, Xuehang and Lu, Zhiyong and Hope, Tom and Wang, Qingyun},
  journal={arXiv preprint arXiv:2602.10081},
  url={https://arxiv.org/abs/2602.10081},
  year={2026}
}

---

© 2026 | College of William & Mary, Allen Institute for AI, NIH

🥰 Thanks much for your interest!