Core positioning of the ReCall framework
ReCall is an open source framework based on reinforcement learning, specialized in training large language models (LLMs) to implement tool invocation and multi-step inference. Its most important feature is to completely get rid of the dependence on supervised data, through the mechanism of autonomous learning to allow the model to master the ability to use a combination of complex tools.
Three technical features
- unsupervised learning: Employing the verl framework of reinforcement learning algorithms to automatically optimize tool invocation strategies through environmental interactions
- Toolset capacity: Supports dynamic invocation of basic tools such as search and calculator, and can handle scenarios where multiple tools are used in tandem tasks
- openness and extensibility: Provide standardized interfaces to allow developers to access any custom tool to form a functionally scalable intelligent body system.
Typical Application Scenarios
Including but not limited to: multi-hop Q&A for cross-document retrieval, enterprise-level automated report generation, complex computational tasks in scientific research scenarios, and other intelligent scenarios that require multi-tool collaboration.
This answer comes from the articleReCall: training large models for tool-call inference via reinforcement learningThe
