Hugging Face published Ecom-RLVE, a training environment for e-commerce AI agents

Hugging Face released Ecom-RLVE — a set of verifiable environments for training conversational AI agents that help customers purchase products in online stores. The project brings reinforcement learning from the world of abstract tasks into real multi-step scenarios: product search, finding substitutes, cart building, returns, and order tracking.

Why Old Benchmarks Aren't Enough

Large language models have long learned to sound convincing, but in e-commerce, that's not enough. A user might ask not simply to "find a charger," but to find a model under $25 with USB-C, two-day delivery, and compatibility with a specific device. For an agent, this is no longer a chat response, but a chain of actions: find the product card, check constraints, select the right variant, get the quantity right, and not invent what doesn't exist in the catalog.

"Fluent speech does not equal task completion."

It's precisely this gap that Ecom-RLVE is built on. The authors develop the idea of RLVE-Gym, where models trained on verifiable tasks with exact rewards, and transfer it to dialogue-based commerce. Instead of subjective evaluation by a human or LLM-as-a-judge, the environment verifies the result with code: did the agent find the right product, correctly select size or variant, create a return for the right item, stay within the step limit.

How the Environment Works

Each episode in Ecom-RLVE is a hidden task, a simulated user, and a set of tools the agent works with. It doesn't just write text—it calls functions, searches the catalog, adds items to the cart, asks clarifying questions, and completes the scenario only when the goal is truly achieved. Eight types of situations form the basis: from product discovery and product substitution to bundle planning, policy QA, order tracking, and multi-intent journey.

The reward is assembled from multiple components so the model learns not just to "appear helpful," but to see the task through to completion:

• reward for correct task completion
• bonus for fewer steps and reduced tool calls
• penalty for hallucinations, such as non-existent SKUs or variants
• hard failure for invalid actions and format violations

Adaptive difficulty is separately important. Instead of fixed easy/medium/hard levels, the environment introduces a complexity number d that controls 12 axes at once: number of constraints, missing details, similar products, typos, out-of-stock items, intention changes mid-dialog, and other obstacles. This makes it possible to build curriculum learning without manual annotation and not keep the model too long on tasks that have become trivial.

Where the Model Fails

The paper details the Cart Building scenario, where the agent must assemble a cart of multiple products with exact variants and quantities. To prevent rote template learning, developers synthesize variants on the fly: for electronics it might be connector type, for clothing — size, for kitchen goods — material or color. Because of this, the model must not just "recognize the product," but actually link the user request to the right modification within the catalog.

On this environment, the team trained Qwen 3 8B using the DAPO method over 300 steps on the C1 collection, and the benchmark itself provides C2, C4, and C8 modes for training on two, four, and eight environments. The catalog was scaled to two million products through FAISS indexing and gte-modernbert-base embeddings, and the user simulator was built on Qwen3.5-9.7B. As a result, the agent was able to consistently progress to more complex episodes, and the errors themselves became clearly visible: the model might select the right product but miss on variant, forget one order item, or claim a needed version doesn't exist when it saw it steps earlier.

What It Means

For the AI-shopping market, this is an important shift: competition can now be not about how smoothly the bot talks, but how reliably it completes the purchase task. If such open environments take hold, the industry will gain a more honest way to train and compare e-commerce agents — by actual action quality, not by dialog impression.