ReCall

One-Minute Overview#

ReCall is an innovative reinforcement learning framework that trains large language models to reason through intelligent tool calling, enabling them to solve complex problems without requiring supervised data on tool use trajectories. It empowers LLMs to act as intelligent agents that proactively use and combine various tools like OpenAI o3, providing a viable approach toward building general-purpose AI agents. As the successor to ReSearch, ReCall extends beyond single search tools to support reasoning with any user-defined tools.

Core Value: Enhances LLM reasoning capabilities through reinforcement learning without requiring supervised tool use data

Quick Start#

Installation Difficulty: High - Requires Python 3.10, GPU resources, and installation of multiple complex dependencies

conda create -n re-call python==3.10
conda activate re-call
git clone https://github.com/Agent-RL/ReCall.git
cd ReCall
pip3 install -e .
pip3 install flash-attn --no-build-isolation
conda install -c pytorch -c nvidia faiss-gpu=1.8.0

Is this suitable for me?

• ✅ Enhanced LLM reasoning: Need LLMs to proactively use tools to solve complex problems
• ✅ Multi-hop QA systems: When you need to handle questions requiring multiple retrieval and reasoning steps
• ✅ Reinforcement learning researchers: When you need to train LLMs for tool use capabilities
• ❌ Resource-constrained environments: Requires multiple GPUs and substantial computational resources
• ❌ Plug-and-play needs: The project requires configuration of multiple service components

Core Capabilities#

1. Tool Call Reasoning - Without Supervised Data#

• Trains LLMs to autonomously use and combine tools through reinforcement learning, without requiring predefined tool use trajectories or supervised data on reasoning steps Actual Value: Reduces model training costs while enhancing problem-solving capabilities for complex tasks

2. Multi-hop QA Processing#

• Handles complex questions requiring multiple information retrieval and reasoning steps through Wikipedia search and reasoning tool combinations Actual Value: Improves accuracy and complex problem handling capabilities of question-answering systems

3. Secure Sandbox Execution Environment#

• Provides a sandbox environment for Python code execution to ensure the safety of tool calls Actual Value: Executes untrusted code in a secure environment, protecting system integrity

4. Synthetic Data Generation#

• Generates synthetic data with diverse environments and complex multi-step tasks, helping models develop sophisticated tool-based reasoning capabilities Actual Value: Expands model training data and improves generalization abilities

Technical Stack & Integration#

Development Language: Python Key Dependencies: verl (0.3.0), vllm (0.8.4), SGLang, FastAPI, FlashRAG Integration Method: SDK / Library

Maintenance Status#

• Development Activity: High - First version released in April 2025 with continuous recent updates
• Recent Updates: Recently released trained models and complete implementation
• Community Response: Active - Multiple supporting tools and service implementations available

Documentation & Learning Resources#

• Documentation Quality: Comprehensive - Includes complete documentation for installation, quick start, training, inference, and evaluation
• Official Documentation: README.md (included in GitHub repository)
• Example Code: Provides training scripts, inference examples, and evaluation code