Technical Challenges and Directions for Large‑Model Applications in E‑commerce

Taobao Group is launching a set of ten large‑model challenges that focus on applying foundation models to e‑commerce scenarios. The goal is to invite researchers to address key technical problems and advance the capabilities of AI in commercial settings.

1. Professional‑domain pre‑training models – E‑commerce requires models that can provide accurate, domain‑specific answers. Challenges include constructing high‑quality professional data, preventing catastrophic forgetting, and balancing domain expertise with general language ability.

2. Complex task decomposition and reasoning QA – User queries often hide multiple logical steps (e.g., holiday‑related gifts). Models must incorporate consumer‑level common sense and perform multi‑step reasoning to bridge the gap between natural language expressions and product specifications.

3. Long‑text handling – Multi‑turn retrieval‑augmented dialogues and tool‑calling scenarios generate very long contexts. The challenge is to extend context windows efficiently without degrading model performance.

4. Knowledge hallucination mitigation – Large models may generate confident but incorrect facts. Two research directions are proposed: improving factual memorisation and enabling the model to recognise its own knowledge limits and refuse answering when uncertain.

5. Tool‑use path decision – In e‑commerce, models must select and orchestrate specialized tools (e.g., product search, decision‑factor generators) and integrate heterogeneous tool outputs (documents, tables, images) into coherent responses.

6. Retrieval‑augmented information utilization – When retrieval returns noisy or partially irrelevant data, models need to summarise relevant facts, reject incorrect information, and fuse retrieved knowledge with user intent.

7. Tool calls under fuzzy intent – Users often express vague goals. The model should learn from massive behavioural data to infer the appropriate tool‑calling strategy without a fixed intent mapping.

8. Multi‑objective e‑commerce RLHF – Reinforcement learning from human feedback must be scaled to millions of domain‑specific reward examples and optimise for professionalism, accuracy, coverage, and depth simultaneously.

9. Generative query understanding – Improving query rewriting for search engines so that rewritten queries increase product recall, especially for long‑tail and ambiguous inputs.

10. Cognitive recommendation based on model knowledge – Leveraging large‑model world knowledge to produce “surprising yet sensible” recommendations that go beyond traditional behaviour‑driven methods.

The document concludes with a bibliography of recent surveys and papers on augmented language models, autonomous agents, tool learning, and RLHF.