SciAgentGym: Benchmarking Multi-Step Scientific Tool-use in LLM Agents
This addresses a critical bottleneck for researchers and developers in building autonomous scientific agents by providing a scalable benchmark and method to improve multi-step tool-use, though it is incremental as it builds on existing agent frameworks.
The paper tackles the problem of LLM agents struggling with multi-step scientific tool-use by introducing SciAgentGym, a benchmark with 1,780 tools across four disciplines, and finds that state-of-the-art models like GPT-5 see success rates drop from 60.6% to 30.9% as interaction complexity increases, but their proposed SciForge method enables a smaller model (SciAgent-8B) to outperform a larger one (Qwen3-VL-235B-Instruct) and show cross-domain transfer.
Scientific reasoning inherently demands integrating sophisticated toolkits to navigate domain-specific knowledge. Yet, current benchmarks largely overlook agents' ability to orchestrate tools for such rigorous workflows. To bridge this gap, we introduce SciAgentGym, a scalable interactive environment featuring 1,780 domain-specific tools across four natural science disciplines, supported by a robust execution infrastructure. Complementing this, we present SciAgentBench, a tiered evaluation suite designed to stress-test agentic capabilities from elementary actions to long-horizon workflows. Our evaluation identifies a critical bottleneck: state-of-the-art models struggle with complex scientific tool-use. Even for a leading model like GPT-5, success rates drop sharply from 60.6% to 30.9% as interaction horizons extend, primarily due to failures in multi-step workflow execution. To address this, we propose SciForge, a data synthesis method that models the tool action space as a dependency graph to generate logic-aware training trajectories. By fine-tuning on these trajectories, our SciAgent-8B outperforms the significantly larger Qwen3-VL-235B-Instruct while exhibiting positive cross-domain transfer of scientific tool-use capabilities. These results underscore the promising potential of next-generation autonomous scientific agents.