[Paper] Anagent For Enhancing Scientific Table & Figure Analysis

Source: arXiv - 2602.10081v1

Overview

The paper introduces AnaBench, a new benchmark that captures the real‑world difficulty of interpreting scientific tables and figures across nine domains, and Anagent, a multi‑agent AI system designed to tackle those challenges. By breaking down complex visual‑textual tasks into coordinated subtasks, the authors demonstrate sizable gains over existing models, highlighting a promising direction for AI‑assisted scientific literature analysis.

Key Contributions

• AnaBench dataset: 63,178 annotated instances covering tables and figures, labeled along seven complexity dimensions (e.g., multi‑step reasoning, cross‑modal linking).
• Anagent framework: A four‑agent architecture (Planner, Expert, Solver, Critic) that decomposes, retrieves, synthesizes, and refines analysis of scientific visuals.
• Modular training pipeline: Combines supervised fine‑tuning with specialized reinforcement learning for each agent, preserving both individual competence and collaborative behavior.
• Extensive evaluation: Tests on 170 sub‑domains show up to 13.4 % improvement without any fine‑tuning and 42.1 % improvement after fine‑tuning, establishing new state‑of‑the‑art performance on AnaBench.
• Open‑source release: Benchmark, code, and pretrained agents are publicly available, encouraging reproducibility and further research.

Methodology

1. Task Decomposition (Planner) – The Planner receives a user query (e.g., “What trend does Figure 3 show about catalyst efficiency?”) and splits it into concrete subtasks such as “extract caption,” “locate axis labels,” and “compare values across rows.”
2. Targeted Retrieval (Expert) – Each subtask is handed to the Expert, which invokes domain‑specific tools (OCR, table parsers, figure caption generators) to pull the exact pieces of information needed.
3. Synthesis (Solver) – The Solver stitches the retrieved snippets together, using a language model to produce a coherent, citation‑rich answer that respects scientific conventions.
4. Iterative Refinement (Critic) – The Critic evaluates the draft across five dimensions— factual correctness, completeness, relevance, clarity, and citation quality— and feeds back improvement instructions to the Solver. This loop repeats until the answer meets a quality threshold.

Training proceeds in two stages:

• Supervised fine‑tuning on the AnaBench annotations to teach each agent its core skill.
• Reinforcement learning where the Critic’s quality scores serve as rewards, encouraging the Planner‑Expert‑Solver team to generate higher‑scoring outputs.

Results & Findings

• Baseline vs. Anagent (no fine‑tuning): Average accuracy on AnaBench rises from 58.2 % to 71.6 % (+13.4 %).
• With domain‑specific fine‑tuning: Accuracy climbs to 82.3 % (+42.1 % over baseline).
• Ablation studies reveal that removing the Planner drops performance by ~9 %, while omitting the Critic reduces final answer quality by ~15 % (measured via human evaluation).
• Cross‑domain robustness: Gains persist across all nine scientific fields, indicating the framework’s ability to generalize beyond a single discipline.

Practical Implications

• Literature review automation: Researchers can query large corpora of papers and receive concise, citation‑backed explanations of data presented in tables/figures, dramatically cutting manual extraction time.
• Intelligent scientific assistants: Development of IDE‑style plugins (e.g., for Overleaf or Jupyter) that surface contextual insights from embedded figures as developers write papers or reports.
• Data‑driven product documentation: Companies can automatically generate user‑friendly summaries of technical specifications that are often hidden in dense tables.
• Regulatory compliance: Automated extraction of quantitative evidence from scientific reports can streamline audit trails for health, safety, or environmental certifications.

Limitations & Future Work

• Tool dependence: The Expert’s performance hinges on the quality of OCR and table‑parsing utilities; errors in low‑resolution figures still propagate.
• Long‑context handling: While the Planner can split tasks, extremely long multi‑figure narratives occasionally exceed the language model’s context window, leading to incomplete reasoning.
• Domain adaptation cost: Fine‑tuning yields the biggest gains but requires labeled data for each new sub‑domain, which may be scarce.
• Future directions suggested by the authors include tighter integration of visual grounding models, scaling the Critic’s evaluation to incorporate domain‑expert feedback, and exploring few‑shot adaptation techniques to reduce annotation overhead.

Authors

• Xuehang Guo
• Zhiyong Lu
• Tom Hope
• Qingyun Wang

Paper Information

• arXiv ID: 2602.10081v1
• Categories: cs.CL, cs.AI
• Published: February 10, 2026
• PDF: Download PDF