Paper 2509.01055
VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use
- Published
- Sep 2025
- Research lab
- Independent
- Citations
- 81
- GitHub
- 1K stars
01 In brief
Summary
VERLTOOL is a unified, modular framework for Agentic Reinforcement Learning with Tool use (ARLT), addressing fragmentation, synchronous execution bottlenecks, and limited extensibility in existing ARLT codebases.
It builds on VeRL, providing upstream alignment, a standardized tool server API supporting diverse tools (code execution, search, SQL, vision, bash), and asynchronous rollouts that achieve nearly 2x speedup.
The framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens, extending single-turn RLVR.
Evaluations across six domains (math, knowledge QA, SQL, visual reasoning, web search, software engineering) show competitive performance compared to specialized systems.
Key design choices include masking off-policy observation tokens during GRPO optimization, separate tokenization of actions and observations to avoid tokenization inconsistencies, and a plugin architecture where new tools require only lightweight Python definitions.
Training dynamics reveal task-specific tool usage patterns, with models learning to reduce unnecessary tool calls in SQL tasks but increasing search usage in deep search tasks.
The framework also supports multi-threading and Ray-based parallel execution for scalable tool serving.
02 From the paper
Abstract
Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2$\times$ speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research. Our code is open-sourced at https://github.com/TIGER-AI-Lab/verl-tool.