ReCall

Why we included this project

Most teams training agentic LLMs find that getting a model to call tools like search or code execution requires hand-labeled trajectories, which is slow and expensive. ReCall avoids that by training purely through reinforcement learning, letting the model learn to reason with and combine arbitrary user-defined tools without any supervised tool-use data. The repo builds on verl and vLLM and includes the training code, a Python sandbox for running tool calls, a Wikipedia retriever service, and prepared datasets like SynTool and MuSiQue, so you can reproduce the pipeline or swap in your own tools. The blog and arXiv paper explain the design, which is useful if you want to understand the approach before running it. It is clearly a research codebase rather than a finished product, so expect to extend it rather than deploy it as-is.