New Research Lets Your Video Feed Get Smarter for Less
Better recommendations without the giant energy bills that usually come with them.
Every time you open a short-video app or a shopping site, a recommendation model decides what to show you next. These systems get better by getting bigger, but training a bigger version from scratch takes enormous amounts of data, time, and computing power — and then it costs even more to run. A team of researchers posted a paper describing a shortcut called Inherit4Rec: instead of rebuilding the model, the new, larger version inherits what the old one already learned, the way you'd upgrade the engine in a car rather than buy a whole new car.
The paper tackles two kinds of upgrades. The first keeps the model "dense," meaning one big brain that handles every request. The second converts it into a "mixture of experts," which is like swapping one generalist for a team of specialists — only the few relevant specialists wake up for each request, so it runs cheaper. The tricky part is doing this without the model forgetting its old skills, or letting some specialists get overloaded while others sit idle.
The team tested both upgrades on KuaiRand-1K, a public short-video dataset, and on real data from an industrial short-video service. Both versions consistently beat the older inheritance methods they compared against. In plain terms: the smarter model didn't get dumber when it grew, and the cheaper version kept most of the quality.
So what does this mean for you? If these techniques reach real products, feeds could improve faster and cost less to run — and cheaper recommendations are one reason apps can afford to serve more people, more personally. The honest catch: this is a research paper, not a shipped feature. It was tested on datasets, not on your actual app, and gains in the lab don't always survive the trip to your phone.
- Recommendation AI usually has to be retrained from scratch to get bigger — this method lets it inherit what it already knows instead.
- One version turns a single big model into a team of specialists, so only the useful parts run for each request, saving computing power.
- It beat older methods on a public short-video dataset and on a real industrial video service's data, but it's research — not a feature you can use today.
Why It Matters
Cheaper, faster recommendation AI could mean better feeds for you and lower costs for the apps serving them.