SciToolAgent-Evo: An Ontology-Aware Self-Evolving Agent for Open-World Scientific Tool Acquisition

2026-07-30

AI Summary

This work addresses the challenge that large language model agents in open-world scientific research struggle to adapt to dynamic tool requirements due to static tool spaces. To overcome this limitation, the authors propose a self-evolving agent endowed with ontological awareness. The approach constructs an ontologized tool graph and accumulates skill memory through contrastive trajectory knowledge distillation. During inference, it dynamically balances exploration and exploitation via a LinUCB multi-armed bandit mechanism and incrementally enriches the scientific ontology with newly acquired tools for seamless integration. This study is the first to integrate ontological awareness with a self-evolution mechanism, enabling continuous acquisition, comprehension, and incorporation of novel tools. Evaluated on the newly introduced OpenSciToolBench benchmark—comprising 900 real-world tasks spanning four difficulty levels—the method achieves state-of-the-art performance, significantly enhancing robustness and generalization.

Abstract

Large language model (LLM) agents have been increasingly adopted in scientific research for organizing and invoking specialized computational tools. However, their reliance on predefined tool spaces with static semantics limits their applicability to open-world scientific workflows, where tool requirements, capabilities, and boundaries evolve dynamically. To this end, we propose SciToolAgent-Evo, an ontology-aware self-evolving agent for open-world scientific tool acquisition. Driven by an evolving memory of skills, experiences, and an ontologized tool graph, it distills generalizable knowledge from contrastive trajectories during accumulation, whereas during inference, it formulates active requests and utilizes a LinUCB-based bandit gate to dynamically balance exploration and exploitation. Once a novel tool is acquired, its scientific ontology is completed online for seamless integration into the known graph. Moreover, we introduce OpenSciToolBench, a benchmark containing 900 realistic tasks across four difficulty levels. Extensive evaluations show that SciToolAgent-Evo achieves state-of-the-art performance, validating its robustness and generalization.

Problem

• open-world scientific workflows
• tool acquisition
• dynamic tool boundaries
• ontology-aware
• LLM agents

Innovation

• ontology-aware
• self-evolving agent
• open-world tool acquisition
• LinUCB bandit
• tool graph