# Pinterest’s PinGER retrieves Pins with fewer input tokens

> A new retrieval system reported in an arXiv paper uses compact histories and controlled generation. Pinterest says it improved saves in Related Pins.

Oossa · 2026-10-27 · https://oossa.com/en/pinterest-s-pinger-retrieves-pins-with-fewer-input-tokens

A team of researchers has introduced PinGER, a system Pinterest uses to find recommendations for its Related Pins feature. The system replaces a nearest-neighbor search with generative retrieval: instead of searching for a close match among item representations, a model generates an item’s semantic ID, a sequence of tokens used to identify it.

The paper focuses on a trade-off in earlier systems. Those systems use several tokens to represent each item in a person’s activity history. With a fixed context window—the amount of input the model can consider at once—that means the model can fit fewer items than if each one used a single embedding. PinGER’s authors say the system keeps semantic IDs for generated results while making those input histories more compact.

## How does PinGER keep histories compact?

PinGER interleaves semantic-ID tokens with dense tokens, which provide a compact representation of an item. Its attention mask lets later items attend only to the dense tokens. The authors say this lets the model learn from every semantic-ID position during training, but serve one token per item at inference while computing the same function. That is the core change intended to let a fixed context include more of a user’s history.

The system also lets Pinterest steer retrieval toward different goals. The authors describe goal-conditioned outputs, reinforcement learning and a penalty that discourages repeated prefixes in generated IDs. Together, these are meant to support surface-specific aims and a more varied candidate set. The paper’s abstract does not explain how those goals are configured for individual users.

## What changed in Related Pins?

The authors report that PinGER is deployed in Pinterest’s Related Pins and increased saves of fresh Pins by 3.60% and overall saves by 0.62%. They also report a 0.7% reduction in the surface’s end-to-end latency and say PinGER eliminated the cost of the nearest-neighbor search it replaced. These are results reported by the system’s authors; the abstract does not give a measurement period or traffic figures.

PinGER’s serving setup is designed for wide-beam generation, which means exploring many possible outputs, and restricts decoding to valid item identifiers. It then resolves those candidates using an index stored on the host machine. The paper presents this combination of compact inputs, controlled generation and serving infrastructure as part of a system already running in a recommendation product, rather than only a lab test.

## The facts

- PinGER is deployed in Pinterest’s Related Pins feature, according to its authors.
- The authors report a 3.60% increase in saves of fresh Pins and a 0.62% increase in overall saves.
- The authors report a 0.7% reduction in Related Pins’ end-to-end latency.
- PinGER replaces the nearest-neighbor search and uses an index stored on the host machine to resolve candidates.

## Why it matters

For people browsing Related Pins, the reported increase in saves suggests the recommendations may lead to more Pins being saved, while the measured latency change is small. The abstract does not say how long the results were measured or how they vary across users, so it cannot show whether an individual will notice a difference.

## Sources & references

1. [PinGER: Token-Efficient Generative Retrieval with Semantic IDs](https://arxiv.org/abs/2610.33435) – arXiv cs.IR, 2026-10-27

Last updated: 2026-10-27