From the source
Ant Ling introduced AgenticLearning, an open-source agent training paradigm for multi-turn interactions with environments.
The project includes two frameworks: RAG-R1 for deep search and reasoning via multi-query parallelism, and FunReason for enhancing function calling capabilities through automated data refinement and a Self-Refinement Multiscale Loss approach.
FunReason achieves GPT-4o-comparable performance on BFCL and mitigates catastrophic forgetting on HumanEval and MBPP.
