Improving RL Outcomes with Exa

Improving RL Outcomes with Exa

Teams training search agents with RL are under pressure to hit benchmark targets while keeping compute and inference costs down. Most treat the search tool as a fixed part of the setup, but the backend an agent trains on shapes how fast it learns, how many calls it makes, and how well it performs once deployed.

In this session, the Exa research team shares what they've learned working with frontier labs that train on Exa instead of traditional web search.

What you'll learn

• How Exa-trained agents compare to SERP-trained agents on pass@k, token count, and search calls per rollout
• Where the training gains come from: search results that contain the answer in fewer calls
• How to keep rewards honest after the agent starts gaming your grader

Who should attend

• ML engineers and researchers post-training agents with RL
• Teams building deep research or search agents on live web data
• Anyone choosing a retrieval backend for a training pipeline and wondering whether it matters