WooooDyy/AgentGym
A framework and benchmark for training, evolving, and evaluating LLM-based agents across diverse interactive environments using reinforcement learning.
Velocity · 7d
+1.1
★ / day
Trend
→steady
star history
AgentGym provides a unified platform for developing and benchmarking LLM-based agents across multiple environments including WebArena, MiniWob++, and SciWorld. The framework enables reinforcement learning training of agents in interactive multi-turn decision-making settings. It releases a fine-tuned model (AgentEvol-7B), trajectory datasets (AgentTraj-L), and evaluation benchmarks (AgentEval) to assess agent capabilities across diverse tasks.