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

The paper extends RLVE framework to multi-turn e-commerce conversations, presenting EcomRLVE-GYM for training shopping agents with algorithmically verifiable rewards.

• •Eight verifiable environments span product discovery, cart building, returns, order tracking, policy QA, bundle planning, and multi-intent journeys, eliminating need for human annotation
• •12-axis difficulty curriculum controls constraint complexity, search result quality, stock availability, input noise, and other factors independently
• •Reward combines F1 task completion score, efficiency bonus for fewer turns, and hallucination penalties for unretrieved product recommendations
• •Qwen 3 8B model trained with DAPO over 300 steps using adaptive scheduling that progresses environments by agent success rate
• •