SciToolAgent-Evo: Self-Evolving Agent for Scientific Tool Acquisition

A new research paper presents SciToolAgent-Evo, an innovative large language model (LLM) agent specifically engineered for open-world scientific tool acquisition. Traditional LLM agents in scientific research are often limited by predefined tool spaces and static semantics, which hinder their adaptability to dynamically evolving scientific workflows. SciToolAgent-Evo overcomes these limitations by incorporating an evolving memory of skills and experiences, alongside an ontologized tool graph. This architecture allows the agent to distill generalizable knowledge from its learning trajectories. During inference, it actively formulates requests and employs a LinUCB-based bandit gate to intelligently balance exploration for new tools with exploitation of known ones. When a novel tool is acquired, its scientific ontology is automatically completed online, ensuring seamless integration into the agent's existing knowledge graph. To rigorously evaluate this agent, the researchers also introduce OpenSciToolBench, a new benchmark comprising 900 realistic tasks across four difficulty levels. SciToolAgent-Evo demonstrates state-of-the-art performance on this benchmark, validating its robustness and generalization capabilities.