Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational Agents

Ecom-RLVE is a new framework for training conversational AI agents to handle real-world e-commerce tasks using reinforcement learning. Instead of relying on subjective feedback, agents are optimized for objectively verifiable outcomes, such as whether the correct products were added to a cart or if an order was tracked successfully. The system provides eight distinct environments that simulate shopping scenarios like product discovery, returns, and complex multi-intent conversations. To ensure agents continuously improve, an adaptive difficulty curriculum automatically makes tasks more challenging across 12 different axes as the agent's performance gets better.