#GRPO
2 articles tagged GRPO
Tagged articles
Articles tagged GRPO, newest first.
AI PerformanceAI Agent Tuning: When to Rewrite Prompts and When to Retrain Weights
Prompt tuning works with hundreds of rollouts; weights need thousands of examples. Optimization requires data, a built system and an evaluation metric together. GEPA raised aggregate benchmark scores 12.19 points over a GPT-4.1 mini baseline. Qwen3 on-policy distillation beat RL on AIME'24 with 1,800 GPU hours versus 17,920. Prompts fit API-only models and costly rollouts; weights fit small student models.
AI PerformancePre-Training, Fine-Tuning and RL: Where Frontier LLMs Get Their Quality
Frontier LLM quality comes from three stages: pre-training, fine-tuning and RL. Pre-training yields a raw autocomplete engine that cannot stop or hold a persona. Fine-tuning on formatted conversations adds the assistant persona and stop token. GRPO grades 30 to 100 attempts against the group average, so wrong tries also teach. Compare a base and instruct open model for free in Google Colab to see the stages.
Get new articles by RSS or podcast
Add the RSS URL to your RSS reader or the podcast URL to your podcast app to get every new article as soon as it is published. The podcast reads out the headline and key points.
How to subscribe