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

Summary

Ecom-RLVE extends the RLVE framework from single-turn reasoning puzzles to multi-turn, tool-augmented e-commerce conversations. EcomRLVE-GYM supplies eight environments spanning product discovery, substitution, cart building, returns, order tracking, policy QA, bundle planning, and multi-intent journeys, with procedural problem generation, a 12-axis difficulty curriculum, and algorithmically verifiable rewards. Each episode requires tool use and world-state changes; rewards combine task completion, efficiency, and a penalty for recommending product IDs not retrieved, while malformed outputs and illegal calls fail immediately. Adaptive scheduling advances each environment independently when its success rate is reliable. An initial study trained Qwen 3 8B with DAPO for 300 steps on cart building and reported progressively increasing difficulty reached, consistent with a steady learning signal.

Takeaways

1. 1

   The eight environments cover product discovery, substitution, cart building, return and replacement, order tracking, policy QA, bundle planning, and multi-intent journeys.

2. 2

   Rewards combine goal completion, agent-caused turn efficiency, and a hallucination penalty for recommending product IDs that were not retrieved during the session.

3. 3

   A 300-step DAPO study on Qwen 3 8B using the cart-building environment showed progressive growth in the difficulty reached and a steady learning signal.