Towards Faithful Simulation of Human Shopping Behavior
Jiakai Tang, Yan Mi, Jing Yu, Yang Zhang, See-Kiong Ng, Qi Cao, Fei Sun, Xu Chen, Wen Chen, Jian Wu, Han Zhu, Bo Zheng
cs.IR
2026-08-21
RecVerse uses screenshots, hierarchical memory, and trajectory GRPO; F1 over STA rises from 4.27 to 7.19 and HCO to 32.64, yet humans still pick real users 74% of the time.
Offline evaluation and RL training for e-commerce both need a shopper that browses like a person. Rule simulators have tiny action spaces. LLM agents read item text and miss layout and images. Recent GUI simulators finally take screenshots, yet long sessions still break.
Two concrete failures. Memory: a shopping trip spans many viewports. Dropping history severs cross-page comparison. Concatenating it floods the context and, past a moderate visual-memory length, accuracy falls while token cost keeps rising. Optimization: almost every current simulator imitates the next logged action. Real browsing includes accidental overscrolls that reverse immediately. Step-level fitting learns that noise. Step-level rewards also cannot see whether a whole session clicks too much or too little. RecAgent's add-to-cart rate hits 80.09% and Agent4Rec hits 55.09%, against 15.93% for real users. Intent scores get inflated by over-clicking.
RecVerse is a GUI-grounded simulation agent. It sees page screenshots and emits multi-turn trajectories. The backbone is Qwen3.5-2B with a frozen visual encoder and full-parameter fine-tuning. Imitation learning warms the policy. GRPO then trains on whole sessions.
Memory has three layers. Updates are actions, trained jointly with clicks and scrolls, rather than written by a heuristic.
Reward has three parts. A macro term penalizes relative gaps in item-directed action counts against the real trajectory, pushing back on over-exploration and passivity. A micro term scores hierarchical category overlap on a three-level taxonomy, with purchases weighted above clicks, so exact SKU hits are not required. A format term requires parseable output and an executable action. Imitation data is synthesized by Qwen3.5-397B-A17B, which reconstructs mindset and memory updates from logged actions. Because raw actions are long-tailed, training uses power-law resampling with α=0.3 so the policy does not collapse into scroll-down.
The paper also releases USB: 5,274 real GUI trajectories from a major East Asian marketplace, 69,842 steps, 8 action types, 90,095 items, category depths 41/517/2,256, and 5,222 users. The claim is that USB is the first user-simulation benchmark that jointly offers visual trajectories, diverse actions, user profiles, and an interactive environment for multi-turn agentic RL.
Real-user references: average trajectory length 13.47 steps, CTR 9.08%, add-to-cart 15.93%, conversion 13.54%, item-page-view 29.60%.
RecVerse-GUI with RL: length 16.25, CTR 7.78%, add-to-cart 19.99%, conversion 4.87%, item-page-view 32.86%. Item-level F1 is 7.19, hit rate 10.45, hierarchical category overlap (HCO) 32.64. The strongest GUI baseline, STA, scores F1 4.27, HR 5.92, HCO 23.11. A text-only RecVerse with RL still only reaches HCO 24.38, so screenshots add something.
RecAgent and Agent4Rec post higher category scores while add-to-cart and conversion explode past real users. The paper marks them as behaviorally distorted and excludes them from bolding among comparable methods. In pairwise human judgments, annotators prefer RecVerse to STA in 92% of comparisons, and still prefer real users to RecVerse in 74%. Fleiss' κ is 0.834. At λ=1000 the intent metrics peak, and trajectory-length error grows to 2.78 steps. Scaling to Qwen3.5-4B lifts category metrics from the low 30s to above 40% and moves conversion closer to real users, while item-page-view undershoots. Category matching is a better RL signal than exact SKU hits: F1 moves from 4.90 to 7.19 and HR from 6.51 to 10.45.
Offline recsys evaluation and agentic recommender training need an environment that does not game the metric. The paper splits process-oriented browsing from goal-oriented GUI agents. The former must reproduce comparison, hesitation, and messy scrolling, not complete "buy item X". Treating memory writes as actions, and scoring whole sessions instead of next-step correctness, transfers to other long-horizon GUI simulators.
Do not read "beats STA" as production-ready. Item-level F1 of 7.19 means exact hits are still rare. Conversion at 4.87% versus 13.54% for real users is a large miss. The simulator looks more like shopping than prior GUI agents. It is not a stand-in for live traffic. USB claims interactivity, from a single marketplace. External reproducibility is unproven.
The authors already report a clear human gap versus real users and call for deeper personalization. Conversion stays misaligned. Raising λ improves intent and lengthens trajectories, so the two rewards fight. The micro reward uses categories from the reference trajectory, not an independent intent label, which leaks session tags into the score. Training and inference cap at 20 steps. Real users average about 13, but the introduction's "dozens of pages" setting is not fully trained. The visual encoder is frozen. The main table is 2B; 4B results live in a figure. USB comes from an unnamed East Asian marketplace, and the paper does not make the interactive environment's public availability checkable.