SciToolAgent-Evo: Self-Evolving AI for Open-World Scientific Tool Acquisition
A recent study presents SciToolAgent-Evo, an ontology-aware self-evolving agent tailored for the acquisition of scientific tools in open-world environments. This agent overcomes the shortcomings of large language model (LLM) agents, which depend on fixed tool spaces with unchanging semantics that fall short in dynamic scientific settings. SciToolAgent-Evo leverages an evolving memory of experiences and skills, along with an ontologized tool graph, to extract generalizable insights from contrastive trajectories. During inference, it generates active requests and utilizes a LinUCB-based bandit gate to strike a balance between exploration and exploitation. When a new tool is obtained, its scientific ontology is updated online for smooth integration into the existing graph. Additionally, the paper introduces OpenS, a benchmark for open-world scientific tool acquisition. This research can be found on arXiv with the identifier 2607.28692.
Key facts
- SciToolAgent-Evo is an ontology-aware self-evolving agent for open-world scientific tool acquisition.
- It addresses the limitation of LLM agents relying on predefined tool spaces with static semantics.
- The agent uses an evolving memory of skills, experiences, and an ontologized tool graph.
- It distills generalizable knowledge from contrastive trajectories during accumulation.
- During inference, it formulates active requests and uses a LinUCB-based bandit gate to balance exploration and exploitation.
- Novel tools are integrated by completing their scientific ontology online.
- The paper introduces OpenS, a benchmark for open-world scientific tool acquisition.
- The research is available on arXiv under identifier 2607.28692.
Entities
Institutions
- arXiv