#Grpo

A controlled GRPO-vs-rejection-sampling study, a hidden vocabulary-collapse signal in DPO, and new LoRA internals and infra tricks.

GRPO trains an LLM by scoring groups of sampled answers against each other, with no critic model. How it works, memory math for 8B, and how to run it.

Fine-tuning Qwen3 for advanced math reasoning using GRPO: boosting precision, structure, and problem-solving accuracy post-training.

...or how to make Qwen learn query optimization via agentic reinforcement learning

RRCM uses GRPO to learn when to retrieve evidence for LLM recommendation, outperforming fixed-context baselines.