![\"verifiable_signals_dark\"](\"https://cdn-uploads.huggingface.co/production/uploads/6893dd21467f7d2f5f358a95/dA0i6ZB3JDG-rqQtLRCy0.png\")
\n\nTL;DR ā We extend the RLVE framework from single-turn reasoning puzzles to multi-turn, tool-augmented e-commerce conversations. EcomRLVE-GYM provides 8 verifiable environments ā product discovery, substitution, cart building, returns, order tracking, policy QA, bundle planning, and multi-intent journeys ā each with procedural problem generation, a 12-axis difficulty curriculum, and algorithmically verifiable rewards. We train a Qwen 3 8B model with DAPO over 300 steps and present early results demonstrating that environment scaling and adaptive difficulty transfer to agentic, real-world task completion.
\n
This project originated in the Pytorch OpenEnv Hackathon and is still evolving, follow us for updates š„
\nLarge language models can hold fluent conversations, yet deploying them as shopping assistants reveals a persistent gap: fluency ā task completion. A customer who asks "find me a USB-C charger under $25 that ships in two days" needs an agent that invokes the right catalog search, filters on three hard constraints, avoids hallucinating product IDs it never retrieved, and handles follow-ups when the top result goes out of stock.
\nSupervised fine-tuning can teach surface-level tool use from demonstrations, but it cannot scale to the combinatorial space of constraint configurations, partial-information dialogues, and multi-step transactional workflows that real e-commerce demands.
\nReinforcement learning with verifiable rewards (RLVR) offers an alternative: the agent optimises for outcomes ā did the products satisfy the constraints? Was the cart correct? Was the return initiated for the right order line? The challenge is constructing reward functions that are both verifiable (no LLM-as-a-judge subjectivity) and adaptive (difficulty that grows with the policy's capability).
\nRLVE-Gym provides 400 environments for sorting, multiplication, Sudoku, and other algorithmic-reasoning tasks; however, those are all single-turn, text-in / text-out puzzles ā extending to agentic domains was left as future work.
\nEcomRLVE-GYM fills that gap: we stay in the verifiable regime (e-commerce outcomes can be checked algorithmically) while extending to multi-turn, tool-augmented, agentic conversations ā environments where the agent must act (call tools, modify world state) rather than merely reason (produce a text answer) and compensates for the deficiency of the search system.
\nEcomRLVE-GYM transforms customer-service outcomes structurally verifiable:
\n
Every signal above can be evaluated by a program with access to the hidden ground-truth goal. No human annotation or LLM-as-a-judge is needed.
\nBefore we explain the framework, here is what a single EcomRLVE episode looks like at difficulty d = 4. The environment generates a hidden goal, a simulated user opens the chat, and the agent must use tools to satisfy the request. Every action is verified algorithmically ā no LLM judge required.
![\"Sample](\"https://cdn-uploads.huggingface.co/production/uploads/6893dd21467f7d2f5f358a95/qDXS-CPl8DT4JN6Uq6nrt.png\")
The reward is fully computed by code: F1 over (product, variant, qty) tuples, an efficiency bonus for finishing in fewer turns, and a hallucination check that every recommended product ID was actually retrieved. If the agent had picked the Lightning variant instead of USB-C, the simulated user would have corrected it mid-dialogue ā and the F1 would have dropped.
Each environment covers a distinct real-world shopping scenario. The agent must complete the task using tools (catalog search, cart operations, order lookups, policy queries) and is scored by a program ā not a human or another LLM.
\n| Environment | \nWhat the agent must do | \n
|---|---|
| Product Discovery | \nFind products that satisfy all the user's constraints | \n
| Substitution | \nAn item is out of stock ā find a similar, compatible alternative | \n
| Cart Building | \nAdd the exact products, variants, and quantities the user asked for | \n
| Return + Replacement | \nIdentify the right order line, initiate a return, suggest a replacement | \n
| Order Tracking | \nResolve which order the user means and report its current status | \n
| Policy QA | \nAnswer a deterministic question about store policy (return window, shipping rules, etc.) | \n
| Bundle Planning | \nRecommend a complete shopping list for a project within a budget | \n
| Multi-Intent Journey | \nHandle a conversation that chains 2ā5 of the above tasks in sequence | \n
Every environment uses the same three-part reward signal:
\nInvalid outputs (malformed JSON, illegal tool calls) trigger an immediate failure score, creating a strong incentive for well-formed responses from step one.
\nA single difficulty number d controls 12 independent aspects of a task simultaneously. This is important because e-commerce conversations are hard in many different ways at once ā not just along one dimension.
![\"Screenshot](\"https://cdn-uploads.huggingface.co/production/uploads/6893dd21467f7d2f5f358a95/SALZRvBC6TP1HG1ZxqWsh.png\")
Here are four representative difficulty axes:
\n| What changes | \nEasy (d = 0) | \nMedium (d = 6) | \nHard (d = 12) | \n
|---|---|---|---|
| How many constraints the user has | \n2 | \n5 | \n8 | \n
| How often the user omits a constraint | \n5% | \n70% | \n~80% | \n
| Fraction of search results that are distractors | \n0% | \n12% | \n24% | \n
| Items that go out of stock mid-conversation | \n0% | \n30% | \n50% | \n
The other eight axes cover turn budget, input noise (typos, slang), context switches, retrieval depth, order-history size, policy complexity, and tool budget. The full breakdown is in the technical report.
\nAdaptive scheduling. Each environment tracks the agent's success rate independently and only advances to harder problems once the agent is passing the current level reliably. This keeps every environment training at the agent's capability frontier ā avoiding both "too easy to learn from" and "too hard to make progress on".
\nCart building is a good showcase because it requires the full search ā inspect ā clarify ā act loop, has a binary ground truth, and introduces a challenge absent from most recommendation benchmarks: variant selection.
\nTo succeed, the agent must develop five distinct skills:
\n| Skill | \nWhat it means in practice | \n
|---|---|
| Product Discovery | \nSearch the catalog with well-formed queries to find the right items | \n
| Variant Selection | \nIdentify the correct color, size, or connector type ā not just the right product | \n
| Cart Management | \nAdd items with the exact variant and quantity the user asked for | \n
| Clarification Dialogue | \nAsk the user a focused follow-up when a request is ambiguous (e.g., missing size) | \n
| Multi-Item Orders | \nHandle shopping lists with several different products in a single conversation | \n
The agent uses six tools to accomplish this:
\n| Tool | \nWhat it does | \n
|---|---|
catalog_search | \nSearches the product catalog with a natural-language query | \n
catalog_get_variants | \nReturns available variants (color, size, connector, etc.) for a product | \n
cart_add | \nAdds a product to the cart with a specific variant and quantity | \n
cart_view | \nReads the current cart so the agent can verify it matches the request | \n
user_get_visit_history | \nFetches recently viewed products by user | \n
ask_user | \nSends a clarification question to the customer when a detail is missing | \n
The generator samples 1ā5 target products (scaling in difficulty with d), each potentially requiring a specific variant (USB-C vs Lightning, Matte vs Glossy) and a quantity > 1. The agent must:
catalog.get_variants to see available options(product_id, variant_id, qty) tuples to the cartReal product catalogs have sparse variant data ā many products have none, and those that do typically vary only by colour or size. To create a richer discrimination task, we synthesize variants at episode initialization:
\nconnector_type; clothing ā size; kitchen ā material).{USB-C, Lightning, HDMI}.(product_id, variant_id) ā correct product but wrong variant means the unit is unmatched.| Axis | \nd = 0 | \nd = 3 | \nd = 6 | \nd = 9 | \n
|---|---|---|---|---|
| Distinct items | \n1 | \n2 | \n3 | \n4 | \n
| Variant required | \n21% | \n66% | \n93% | \n99% | \n
| Multi-quantity | \n0% | \n30% | \n50% | \n50% | \n
At d = 0 the agent adds a single product with no variant complexity ā learning the basic catalog.search ā cart.add workflow. At d = 6 it juggles 3 items, nearly all requiring a specific variant, with half needing qty > 1.
The cart must be exactly right ā correct product, correct variant, correct quantity. Partial credit is given for partially correct carts, but a perfect score requires every item to match. If the agent adds the wrong variant, the simulated user corrects it mid-dialogue ("that's the Lightning version, but I need USB-C"), giving the agent a chance to self-correct before the episode ends.
\nTwo real E_CART episodes from a Qwen 3 8B agent. Same environment, same agent ā difficulty alone changes the game.
\n| \n | d = 1 ā 1 item, no variants | \nd = 8 ā 3 items, variants + typos | \n
|---|---|---|
| User message | \n"Shopping list: SonicWave Keypad. Please add everything." | \n"pls add the following to my aart: Eco-Frienly HrenLefa Charger by GreenLeaf, High-Performance AqaPure Air Fryer , Lighrweiht PixelVault Tank Top (3)." ā typos from noise axis | \n
| Turn 1 | \nuser.get_visit_history() ā 4 results. Spots the target immediately. | \nuser.get_visit_history() ā 9 results. Must disambiguate across chargers, fryers, tops. | \n
| Turn 2 | \ncart.add("syn_000162", qty=1) ā Cart: 1 item, $2.74. "I found the SonicWave Keypad." User: "Great, that's the exact one." | \nNo tool calls ā presents 3 product cards. User: "Add the Charger with 100W output, Air Fryer with Charcoal filter, 3 Tank Tops in size XS." | \n
| Turn 3 | \nanswer: { "done": true } ā Done in 3 turns. | \ncatalog.get_variants Ć 3 ā charger: 6 variants (100W, 120W, 30Wā¦), air fryer: 6 (Bamboo, Permanent, Charcoalā¦), tank top: 7 (XS, S, XLā¦). User: "I need three Tank Tops in XS." | \n
| Turn 4 | \n\n | cart.add(charger, "100W", qty=1) ā cart.add(fryer, "Bamboo", qty=1) ā should be Charcoal. cart.add(tank_top, "size_XL", qty=3) ā should be XS. User: "Air Fryer has Bamboo not Charcoal, Tank Top is XL not XS." | \n
| Turn 5 | \n\n | Re-fetches tank top variants, adds 2Ć XS ā doesn't fix the air fryer. User: "Air Fryer still shows Bamboo. I need three XS, not two." | \n
| Turn 6 | \n\n | Re-fetches tank top variants again ā still ignores the air fryer. User: "Please change the Air Fryer variant." | \n
| Turn 7 | \n\n | "I cannot find a Charcoal filter variant." ā It exists; the agent saw it in turn 3 but never selected it. User: "Skip the Air Fryer then." | \n
| Turn 8 | \n\n | answer: { "done": true } ā Cart still wrong: Charger ā, Air Fryer (Bamboo) ā, Tank Top (3ĆXL + 2ĆXS) ā | \n
| Reward | \nr_task = +1.00, r_eff = +0.33, r_hall = 0.00, r_total = +0.80 ā | \nr_task ā 0.00, r_eff = ā0.43, r_hall = 0.00, r_total = ā0.06 ā | \n
| Outcome | \nCart matches goal. 3 turns, 2 effective. | \nWrong variants, wrong quantities, user gave up. 8 turns, 6 effective. | \n
At d=1 the agent solves the task in 3 clean turns. At d=8 it spirals ā picking Bamboo instead of Charcoal, XL instead of XS, never fixing the air fryer despite two user corrections, then hallucinating that the variant doesn't exist. This is exactly the kind of multi-step error cascade that the difficulty curriculum surfaces, and that adaptive training should teach the agent to recover from.
\nA verifiable environment needs a user simulator that behaves realistically. We use Qwen3.5 (9.7B) to generate natural, varied user messages rather than canned templates ā covering everything from typo-filled requests to mid-conversation topic switches.
\nTwo design choices matter for training quality:
\nPreferences match stated constraints. Each simulated user has a hidden set of preferences (price sensitivity, brand loyalty, shipping speed, etc.). These are deliberately biased toward whatever constraints the user communicated ā so if the user said "under $25", the reward function actually cares about price. Without this, an agent could be penalised for correctly following the user's instructions.
\nStrategic omission. The LLM deliberately withholds some constraints from the opening message to force the agent to ask clarifying questions. The system tracks exactly what was and wasn't mentioned, so the agent is never penalised for information it was never given.
\nFollowing RLVE's methodology, we define nested environment collections:
\nC1 ā C2 ā C4 ā C8
\n| Collection | \nEnvironments | \nSkills trained | \n
|---|---|---|
| C1 | \nCart | \nSerarch Query Formulation, Cart Manipulation | \n
| C2 | \n+ Substitution | \nSimilarity reasoning under constraints | \n
| C4 | \n+ Product Discovery, Returns | \nTransactional workflows (Retrieval + recommendation, return initiation) | \n
| C8 | \n+ Status, Policy, Bundle, Journey | \nKnowledge retrieval, planning, compositionality | \n
We hypothesise ā consistent with RLVE's findings ā that C8 agents outperform single-environment specialists, even on the specialist's own task.
\nWe trained Qwen 3 8B with DAPO on C1 (Cart Building) for 300 steps as an initial viability study.
\n| \n | Config | \n
|---|---|
| Base model | \nQwen 3 8B | \n
| Algorithm | \nDAPO (G = 8 rollouts/prompt) | \n
| LR | \n1e-5 | \n
| Catalog | \n2M products, FAISS index with Alibaba-NLP/gte-modernbert-base (768-dim) | \n
| User sim | \nQwen3.5 9.7B | \n
![\"accuracy_levels\"](\"https://cdn-uploads.huggingface.co/production/uploads/6893dd21467f7d2f5f358a95/eWQqFP-PbCJeNsn8klCQZ.png\")
We saw progressive growth in difficulty reached, confirming that adaptive scheduling produces a steady learning signal rather than the saturation (static-low) or starvation (static-high) patterns predicted by the RLVE paper.
\nRun a live episode directly in your browser using the embedded demo below. Here is how to get started:
\nE_CART for cart building or E_PD for product discovery).0 is a simple single-constraint task; 6+ introduces missing information, noisy retrieval, and variant selection.The environments, verifiers, and training configs are all open-source:
\n![\"Models\"](\"https://img.shields.io/badge/%F0%9F%A4%97%20Models-WUFUS-blue\"){kind=icon}
![\"Data\"](\"https://img.shields.io/badge/%F0%9F%A4%97%20Catalog%20Data-Amazebay2M-yellow\"){kind=icon}
![\"Code\"](\"https://img.shields.io/badge/Github-Code-black\"){kind=icon}
![\"Demo\"](\"https://img.shields.io/badge/%F0%9F%A4%97-Space-red\"){kind=icon}
\n\n\nThe 2M-product catalog is on the Hub:
\n\n\n = \n\nZeng, Z., Ivison, H., Wang, Y., et al. (2025). RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments. ICML 2025. arXiv:2511.07317
\nYu, Q., Zhang, Z., Zhu, R., et al. (2025). DAPO: An Open-Source LLM Reinforcement Learning System at Scale. arXiv:2503.14476
\nShao, Z., Wang, P., Zhu, Q., et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300
\nDeepSeek-AI. (2025). DeepSeek-R1: Incentivizing Reasoning in LLMs through Reinforcement Learning. Nature.
\nMeta AI. (2024). Llama 3.1: A Foundation Model for General Intelligence. llama.meta.com
\nQwen Team. (2025). Qwen3 Technical Report. arXiv:2505.09388
\n