SciAgentGym: Benchmarking Multi-Step Scientific Tool-use in LLM Agents

TL;DR

SciAgentGym introduces a comprehensive benchmark and environment for evaluating and improving large language models' ability to perform complex, multi-step scientific tool-use across multiple disciplines, highlighting current limitations and proposing a new training method.

Contribution

The paper presents SciAgentGym and SciAgentBench for benchmarking scientific tool-use, and proposes SciForge, a data synthesis method, to enhance models' multi-step scientific reasoning capabilities.

Findings

State-of-the-art models struggle with complex scientific workflows.

Fine-tuning with logic-aware trajectories improves performance.

Success rates drop significantly with longer interaction horizons.

Abstract

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.