Databricks introduces KARL, a custom knowledge agent trained with reinforcement learning to handle grounded reasoning tasks in enterprise settings. KARL matches frontier model performance while significantly reducing inference cost and latency, trained with only a few thousand GPU hours using entirely synthetic data. It powers Databricks' Agent Bricks Knowledge Assistant and excels at hard-to-verify tasks like document search, fact-finding, and multi-step reasoning. Databricks is also opening the same RL pipeline infrastructure to customers via a Custom RL private preview, allowing enterprises to build their own domain-specific agents optimized for high-volume agentic workloads.
Sort: