Forecasting What Shoppers Will Want Next: eBay's DeepAffinity and the Specialist-vs-Generalist Debate

On September 2, eBay's research team quietly published a paper about a question most personalization systems don't even ask: not what a shopper wants right now, but what they'll want three months from now.

What DeepAffinity Actually Does

Their paper, DeepAffinity, frames aspect-level preference — brand, size, color — as a forecasting problem, not a recommendation one.

Using a buyer's full time-ordered interaction history, the model predicts which aspects they'll gravitate toward next. It's built on small language models with structured prompts and task-specific prediction heads, not one giant general-purpose model.

Specialist Beats Generalist

That's the part worth sitting with. The team found their fine-tuned small models beat standard generative fine-tuning. More surprising: general-purpose open-source LLMs performed poorly on this task without that specialization, even at larger scale.

Old Way vs. New Way

Aspect-level personalization has traditionally leaned on collaborative filtering or session-level clickstream models — reactive, and blind to anything outside the current visit.

The obvious 2026 instinct is to point a big general-purpose LLM at the problem instead. eBay's result cuts against that instinct: a small, purpose-tuned model beat the generalist, at presumably a fraction of the inference cost.

The Open Question

If a small, fine-tuned model beats a general-purpose LLM on a narrow, well-defined prediction problem, how many "wrap it in an LLM" projects on enterprise roadmaps right now are solving the wrong problem — reaching for general capability where specialization would win on both accuracy and cost?

Why This Matters for E-Commerce and Marketplaces

We run into this constantly in our own work: a small model that deeply understands one evolving customer behavior — what someone's about to churn on, what they're about to buy — routinely outperforms a general one bolted on after the fact.

What's your line for when a fine-tuned small model beats reaching for the bigger LLM?