Large Model Application Challenges for E-commerce

Taobao Group releases ten challenge topics for large‑model applications in e‑commerce, inviting researchers and students to solve real‑world problems.

Technical background : Standard LLM chatbots focus on fluency, but e‑commerce requires professional, accurate, and timely answers. General pre‑training data lacks domain‑specific knowledge, and retrieval‑augmented methods are needed.

Key research challenges :

How to collect and construct professional domain training data that covers knowledge gaps of generic corpora.

Efficiently utilize limited domain data without catastrophic forgetting.

Balance domain‑specific expertise with general capabilities in a single model.

Complex task decomposition & reasoning : Enable multi‑step reasoning for queries like “gift for Teacher’s Day on Sep 10”, requiring implicit logical inference.

Long‑text handling : Extend context windows of pre‑trained models via lightweight fine‑tuning or interpolation while maintaining performance.

Knowledge hallucination : Improve factual accuracy and enable the model to admit “I don’t know” when out of its knowledge scope.

Tool‑use path decision : Design agents that select and invoke appropriate tools (e.g., product search, image generation) and integrate heterogeneous tool outputs (documents, tables, images) into coherent answers.

Retrieval‑augmented generation : Summarize and reason over retrieved e‑commerce information, reject incorrect facts, and generate reliable responses.

Fuzzy intent handling : Learn from massive user behavior to infer ambiguous intents and decide optimal tool‑calling strategies.

Multi‑objective RLHF : Build reward models for e‑commerce domains and apply RLHF to optimize professionalism, accuracy, coverage, and depth.

Query understanding : Rewrite user queries to bridge the semantic gap with product titles, improving retrieval recall and relevance.

Cognitive recommendation : Leverage large‑model world knowledge and reasoning to produce “unexpected yet reasonable” recommendations, enhancing discovery beyond traditional behavior‑based methods.

Participants are expected to propose data pipelines, training strategies, and evaluation metrics for each challenge.