Uber Eats wants its feed to read dinner plans in seconds, not hours

Uber’s latest change to its Uber Eats recommendation system is not a cosmetic feed tweak. It changes how the platform reads intent. According to InfoQ, Uber has moved the Home Feed toward near real-time sequence modeling and a new Generative Recommender approach for ranking restaurants and offers.

The old problem in systems like this is not simply knowing whether someone tends to like sushi, burgers, or coffee. The harder problem is knowing what that person wants now, after the last few taps, searches, restaurant views, and abandoned options. If the system sees user behavior with a delay of hours, the recommendation may be technically personalized but operationally stale. Uber is now reducing feature freshness from roughly 24 hours to seconds.

That difference matters. A food delivery recommendation system operates inside a very short decision window: the user is hungry, comparing price, delivery time, cuisine, discounts, and familiar restaurants. A signal created a few seconds ago can outweigh a profile built over months. If the user has just opened several ramen restaurants, the model needs to treat that as context, not as a disposable click.

The second shift is from hand-crafted features to transformer-based sequence modeling. Transformers, introduced in the paper Attention Is All You Need, are useful here because they can model ordered events and relationships between them rather than relying only on aggregates and fixed rules. In the Uber Eats Home Feed, that means the model can better interpret the sequence of actions a user takes during a session.

The third shift is the most interesting one for the industry: Uber is moving from pointwise scoring to listwise GenRec ranking. A pointwise system scores each item almost in isolation. A listwise system evaluates the whole recommendation list and the context in which items compete for position. That matters because a feed is not a lab table. Two strong restaurants do not automatically make a strong list if they are too similar, ignore the user’s current intent, or push more useful choices lower on the screen.

For Uber, this is infrastructure rather than an AI demo. The Home Feed has to work across millions of users, fast-changing demand, local restaurant availability, and shifting marketplace conditions. That is why the move also fits into the broader field of learning to rank, but with the harder constraint of surviving production latency and marketplace scale.

The sober takeaway is this: the update does not prove that generative models magically know what people want to eat. It shows that platforms with enough traffic can move personalization away from slow profiles and toward fast behavioral streams. At that point, recommendation stops being only a question of history and becomes a question of the current moment.