modelscope/AgentEvolver

Implements three self-evolving mechanisms—automatic task generation via self-questioning, experience-guided exploration through self-navigating, and causal credit assignment via self-attributing—integrated into a service-oriented dataflow architecture with standardized environment interfaces and modular LLM/experience management services. Demonstrates significant performance gains on AppWorld and BFCL-v3 benchmarks with 7B/14B models, and extends to multi-agent social reasoning through the Game Arena for interactive evaluation and RL-based training in complex environments like Avalon and Diplomacy.

1,284 stars. Actively maintained with 2 commits in the last 30 days.