Hey guys!🙋 This is FirstLab! 1️⃣FirstLab is an AI dev community with English, Chinese, and Japanese speakers. Join us to stay updated on AI topics, share Dify plugins and MCP configurations, and connect with developers across languages. 💥Break the information gap with us!

We’ve been rethinking how AI agents learn, and we might have just found a new way. Tencent Cloud AI introduces Training Free GRPO, a new way for large models to learn without fine tuning or gradient updates. Traditional reinforcement learning for LLM agents is powerful but costly: thousands of samples, long training cycles, and expensive compute. Training Free GRPO changes that. Instead of updating parameters, the model learns from its own experience, comparing multiple rollouts, learning from what worked, and turning those insights into a reusable semantic memory that guides future reasoning. Why it matters: 🔹No fine-tuning. Faster by magnitudes. 🔹Works with ~100 samples. Highly data efficient. 🔹 Outperforms $10,000+ RL setups for less than $20 🔹> 80 % accuracy on AIME 24/25 and smarter tool use in WebWalkerQA This approach expands model optimization from gradient space to semantic space, enabling low-cost, self-improving agents that evolve through reflection instead of training. This technology will be available in Tencent Cloud ADP, bringing training-free optimization directly into enterprise agentic workflows. It is our sincere hope that this work can bring about meaningful innovations, so we’ve open-sourced the code and have attatched the paper below. We will continue to open-source new research and advancements to push the boundaries of agentic intelligence. Follow us & ⭐ the repo to stay updated. #AI #LLM #Agents #ReinforcementLearning #MachineLearning #PromptEngineering #TencentCloud #TrainingFreeGRPO #TencentCloudADP