Welcome to the AI research bites. This series of short and informative talks showcases cutting-edge research work from ServiceNow AI Research team. The AI Research Bites are open to all, especially those interested in keeping up with the fast-paced AI research community. In this presentation, Massimo Caccia shows how to best allocate training compute between supervised fine-tuning (SFT) on expert demonstrations and reinforcement learning (RL) on the agent’s own trajectories — a trade-off between quality and quantity. The results demonstrate that starting with SFT, then continuing with RL, consistently advances the Pareto front of performance vs compute. Moreover, as the amount of SFT warm-up increases, the optimal RL hyperparameters shift, revealing how prior supervision shapes the efficiency and stability of downstream RL fine-tuning. Paper: https://arxiv.org/abs/2507.04103 Blogpost: https://huggingface.co/blog/ppEmiliano/how-to-train-your-llm-web-agent-a-statistical-diag ServiceNow AI Research team: https://www.servicenow.com/research/

 393      6