This paper introduces a framework that uses Large Language Models (LLMs) to automate scientific visualization workflows from natural language descriptions, enhanced with formal verification to ensure pipeline correctness. The system validates workflows for topological and semantic validity, improving reliability over generic LLM agents.
Key findings
LLM agents can automate scientific visualization workflows with formal correctness guarantees.
The system ensures topologically valid DAGs and semantically valid TDA operations.
Achieved 66.5% success rate on TopoPilot benchmark, outperforming unverified LLMs.
Maintains 95% success rate on adversarial inputs, where baselines degrade to 20%.
Verification adds minimal overhead, practical for interactive visualization workflows.
Limitations & open questions
The framework's performance may degrade with highly complex or ambiguous natural language requests.
Integration with existing tools may require additional้้ and customization.