Training Shopping Bots Too

Extended RLVE from single-turn reasoning problems to multi-turn, tool-augmented e-commerce conversations.

• 8 environments: product discovery, substitution, cart building, returns, order tracking, policy QA, bundle planning, multi-intent journey
• 12-axis difficulty curriculum: jointly controls constraint count, missing information, distractor ratio, inventory changes, turn budget, noise, retrieval depth, policy complexity, and more
• Rewards are all computed to be algorithmically verifiable: task reward, efficiency reward, hallucination penalty
• invalid output such as malformed JSON or illegal tool calls is immediately treated as failure

The core demo is Cart Building (E_CART). It goes through product search, variant selection, cart addition, and clarification dialogue, with correctness verified at the (product_id, variant_id, qty) level. Products were synthesized with natural variations like USB-C / Lightning by category to increase difficulty.

Training results were also presented. Qwen 3 8B was trained for 300 steps with DAPO, and it was reported that environment scaling in the form of C1 ⊂ C2 ⊂ C4 ⊂ C8 along with adaptive scheduling produces a stable training signal. The user simulator used Qwen3.5 9.7B.