Shopping Companion benchmark links long-term preference memory to agentic product selection

Context

Alibaba-affiliated researchers introduced Shopping Companion Bench to test shopping agents on cross-session preference recall, product search, attribute verification, and multi-turn correction over a large real-product catalog.

What changed

The study constructed 1,000 benchmark tasks from a Lazada corpus of 1,298,797 products, combining synthetic preference dialogues with single-product and add-on-deal shopping objectives. Authors also tested tool-wise rewards for memory and product-search operations.

Why it matters

Agentic product discovery can fail after product retrieval if the agent forgets user preferences, skips attribute checks, or mishandles budget and bundle constraints. This provides a benchmark for separating retrieval availability from recommendation validity.

Visibility OS interpretation

Visibility OS interpretation: Product inclusion and citation alone are insufficient shopping success metrics. Preference fit, product-attribute verification, and task completion are separate measurable stages. This is not evidence of commercial AI shopping rankings.

What remains uncertain

Will results replicate with genuine long-term customer histories, other retailer catalogs, or real transactions? Are gains robust across agent models and shopping interfaces?

Evidence claims

Emerging Evidenceintroduces

Shopping Companion Bench comprises 1,000 author-constructed shopping tasks over a corpus of 1,298,797 Lazada products, with single-product and add-on-deal tasks and preference histories checked by two e-commerce experts.

Emerging Evidenceintroduces

The Shopping Companion authors report that their GPT-5 zero-shot baseline achieved under 70% task success on their benchmark, and that tool-wise reward training improved their 4B agent versus their tested training baselines.