Topics · How it works
Sequence modeling
For Instagram creators, sequence modeling is a recommendation approach described in Meta’s research and engineering that uses the order of past interactions to predict content or ads someone may engage with next; Meta does not establish Instagram-specific deployment.12 Meta’s engineering team describes using streams of recent events, with the number of events depending on each stream’s importance and with details such as ad category and interaction time retained.1 In August 2026, Engineering at Meta described combining individual features and ordered behavior so an ads model can learn relationships directly from data rather than manually designed combinations.3 These Meta accounts do not establish user controls for sequence modeling, effects on Instagram creators’ reach, or current creator guidance.123
Learning from ordered interaction histories
Meta research and patents. These may not describe how Instagram works today.
Meta’s engineering team explains that each stream includes a chosen number of recent events, with that number determined by the stream’s importance.1 Meta’s engineering team says each event also retains information about its meaning and context, such as the ad category and when the interaction happened.1 In an August 2026 Engineering at Meta account, Meta’s engineering team describes combining individual features and ordered behavior into a shared representation, allowing the ads model to learn relationships directly from data rather than relying on manually designed combinations.3
Generative recommendations with HSTU
Meta research and patents. These may not describe how Instagram works today.
Meta’s HSTU research extends sequence modeling to two recommendation tasks: retrieving content, meaning selecting candidates, and ranking content, meaning deciding their order.2 Meta’s research calls its approach “Generative Recommenders” and proposes HSTU as a model design for turning an input sequence into those predictions.2 Meta’s research reports that HSTU was 5.3 to 15.2 times faster than the comparison Transformer models using FlashAttention2 when processing sequences of 8,192 items.2 Meta’s research also reports a 12.4% improvement in metrics in live comparisons between system versions, although the supplied passage does not name those metrics.2 Meta’s research says HSTU-based systems were deployed in multiple product areas, but it does not identify these performance figures as Instagram-specific results.2