ReCall
Basic Information
ReCall is a research and engineering repository that provides a training and evaluation framework for teaching large language models to reason by calling external tools via reinforcement learning. The project focuses on enabling LLMs to agentically use and combine arbitrary user-defined tools without supervised trajectories for tool use or stepwise reasoning. The implementation is a successor to an earlier ReSearch project and contains code, data preparation scripts, training recipes, serving utilities and evaluation scripts to reproduce and extend the approach. The repo is intended for developers and researchers who want to train, fine-tune, evaluate and serve models that learn tool-based multi-step reasoning. It bundles a customized reinforcement learning stack, sandboxed tool execution, retriever services and inference wrappers to orchestrate model generation and tool execution in experiments and benchmarks.
Links
Categorization
App Details
-
$$$$ -
$$$$