🤖 machine learning

UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

UniPinRec is the first production system to fully unify retrieval and ranking into a single transformer-based pipeline with shared inputs, training, and serving infrastructure, achieving significant engagement gains and efficiency improvements at Pinterest scale.

Original authors: Hanyu Li, Yi-Ping Hsu, Aditya Mantha, Prabhat Agarwal, Laksh Bhasin, Jialu Wang, Hongtao Lin, Bella Huang, Yaxin Li, Xinyi Li, Chuxi Wang, Kousik Rajesh, Hooshmand Shokri Razaghi, Shunyao Li, Zongyue

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hanyu Li, Yi-Ping Hsu, Aditya Mantha, Prabhat Agarwal, Laksh Bhasin, Jialu Wang, Hongtao Lin, Bella Huang, Yaxin Li, Xinyi Li, Chuxi Wang, Kousik Rajesh, Hooshmand Shokri Razaghi, Shunyao Li, Zongyue Qin, Jaewon Yang, James Li, Dhruvil Deven Badani, Jiajing Xu, Charles Rosenberg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine Pinterest as a massive, bustling library with billions of books (Pins). When you walk in, the library needs to do two things very quickly:

Retrieval: Scan the entire library to find a few thousand books that might interest you.
Ranking: Take those few thousand books, read their summaries carefully, and arrange them in the perfect order for you to see.

The Old Way: Two Separate Librarians

For a long time, Pinterest used two different "librarians" (AI models) to do this.

Librarian A (Retrieval) was a speed demon. They scanned the whole library fast but didn't read the books deeply. They just grabbed a big pile of potential matches.
Librarian B (Ranking) was a careful reader. They took that pile, read every book in detail, and sorted them perfectly.

The Problem: Both librarians were reading the exact same notes about your past behavior (what you clicked, saved, or hid). They were both doing the heavy lifting of "remembering" your history separately. This was like hiring two people to memorize the same phone book just to make two different lists. It wasted money, computer power, and time.

The New Solution: UniPinRec (The Super-Librarian)

The paper introduces UniPinRec, a single "Super-Librarian" that does both jobs at once, using one brain, one set of notes, and one workflow.

Here is how they made it work, using three clever tricks:

1. The "Masked Action" Trick (MAM)

Usually, the "Retrieval" librarian just looks at a list of items you liked. The "Ranking" librarian needs to know what you did with those items (did you click? save? hide?).

The Old Problem: If you mix the "what you did" into the list, it messes up the speed of the Retrieval librarian.
The UniPinRec Fix: They use a "Masking" technique. Imagine the librarian has a list of your history. They cover up (mask) the part that says "what you did" with a sticky note.
- The librarian learns to guess what you did based on the item.
- This allows the librarian to learn both how to find items (Retrieval) and how to judge them (Ranking) without the list getting too long or confusing. It's like learning to drive and park at the same time by practicing with the parking brake on.

2. The "Shared Memory" Trick (KV Cache Reuse)

This is the biggest speed booster.

The Old Problem: When the Ranking librarian started their job, they had to re-read your entire history from scratch to understand who you were.
The UniPinRec Fix: The Retrieval librarian does the hard work of reading your history first. Instead of throwing those notes away, they hand them over to the Ranking librarian.
- The Ranking librarian doesn't re-read the history; they just look at the notes the first librarian already wrote down.
- Analogy: It's like a chef who chops all the vegetables (Retrieval) and then hands the bowl to a sous-chef who just adds the spices (Ranking). The sous-chef doesn't need to chop the vegetables again. This saves a massive amount of time.

3. The "Blended Training" Trick

Instead of training the model to be a "Finder" first and then a "Sorter" later, they trained it to do both simultaneously.

They fed the model examples that had both the "history of what you liked" AND the "list of what you saw but didn't click."
This taught the model to be a better finder because it knew how to judge the items, and a better judge because it knew how to find the right items.

The Results: Faster, Smarter, Cheaper

When they put this Super-Librarian to work in the real Pinterest app, the results were impressive:

Better Recommendations: Users saved more Pins and opened more notifications. The system found better items because the "Finder" and "Sorter" were talking to each other.
Faster Speed: Because they stopped re-reading the user's history, the system became 11% faster overall.
More Capacity: The system could handle 63% more users at the same time without slowing down.

Summary

UniPinRec is like merging two separate departments into one efficient team. By sharing the same "brain" (model), the same "notes" (user history), and the same "workflow," Pinterest saves money on computers, makes the app faster for users, and shows people content they actually want to see. They didn't replace the whole library system; they just made the existing librarians work together perfectly.

Technical Summary: UniPinRec

Problem Statement
Modern industrial recommendation systems, such as those at Pinterest, typically operate as multi-stage funnels where retrieval and ranking are trained as separate models. Despite both stages increasingly relying on large transformer backbones to encode the same user behavior data, this separation creates significant redundancy. The current paradigm duplicates parameters, training compute, and serving costs because retrieval and ranking cannot share learned signals or model weights. While prior work has attempted to unify model architectures (e.g., HSTU, OnePiece), these approaches often fail to unify the full pipeline. They remain fragmented across input formats, training procedures, and serving stacks, or they rely on end-to-end generative methods that replace the entire funnel, making integration with existing candidate sources and operational control difficult.

Methodology
The authors propose UniPinRec, a system that achieves full-stack unification of retrieval and ranking. This approach integrates input formats, model architecture, training, and serving infrastructure into a single, cohesive pipeline deployed within Pinterest's existing production environment. The methodology relies on three core technical innovations:

Masked Action Modeling (MAM):
To enable a single model to serve both retrieval (next-item prediction) and ranking (action prediction) without doubling context length or breaking input compatibility, the authors introduce MAM.
- Architecture: Instead of interleaving action tokens between items (which inflates sequence length), actions are encoded as dense embeddings concatenated with item representations along the feature dimension.
- Masking: During training, actions in the user's history are randomly masked with a probability $p_{mask}$ . At inference, future actions are always masked. This allows the model to learn action prediction objectives (ranking) on the same non-interleaved sequence used for retrieval (next-item prediction).
- Attention Pattern: The system uses a modified causal mask (M-FALCON pattern) where candidate positions attend to all past history but not to each other. This reduces attention complexity from $O((n+k)^2)$ to $O(n^2 + nk)$ and enables efficient KV-cache sharing.
Blended Training and Joint Loss:
The system constructs unified training examples that pair a user's past engagement history with a future feedview impression slate (including negative outcomes).
- Data Infrastructure: A Ray-based in-trainer join merges separate Iceberg tables for user history and feedview impressions, avoiding data duplication and allowing flexible sampling ratios.
- Objective: The model is trained with a joint loss function: $L = L_{item} + L_{action}$ $L = L_{i t e m} + L_{a c t i o n}$ .
  - $L_{item}$ : A sampled softmax loss for next-item prediction (retrieval).
  - $L_{action}$ : A binary cross-entropy loss for predicting specific action types (click, save, hide) on masked positions (ranking).
- This joint optimization allows the item embedder to learn from both tasks simultaneously, improving generalization compared to sequential pre-training and fine-tuning.
Cross-Stage KV-Cache Sharing:
To address serving efficiency, UniPinRec reuses the computation of the user-history encoding across stages.
- Mechanism: The retrieval stage encodes the user history and stores the resulting Key-Value (KV) cache in a pre-allocated GPU memory pool. The ranking stage, running as a separate process, maps this same GPU memory and reuses the cached history without re-encoding it.
- Efficiency: This transforms the ranking stage from a full forward pass ( $O(n^2)$ ) into an incremental decode step ($O(nk)$), where $n$ is history length and $k$ is candidate count.
- Implementation: The system uses NVIDIA Triton for serving, with Faiss for ANN lookup. The Faiss index returns both candidate IDs and their pre-computed embeddings, eliminating the need for a separate embedding fetch during ranking.

Key Contributions

Full-Stack Unification: UniPinRec is the first system to unify retrieval and ranking across inputs, model, training, and serving in a production recommendation system. It replaces the need for two independent models with a single generative decoder.
Efficient Architecture: By introducing MAM and non-interleaved sequences, the system achieves full weight sharing and parameter efficiency without inflating context length.
Serving Optimization: The cross-stage KV-cache sharing mechanism significantly reduces total FLOPs and latency by avoiding duplicate user-history encoding.
Operational Compatibility: The design is a "drop-in" replacement that composes with existing candidate generators (e.g., keywords, trending) and allows for independent A/B testing and rollback of stages.

Results
The system was deployed in Pinterest's core surfaces, yielding the following results:

Offline Performance: UniPinRec matched the production retrieval model's Recall@10 while improving the ranking Hit@3 by +14.8% compared to the dedicated production ranker (TransAct V2 + DCNv2). It also outperformed HSTU and a fine-tuned PinRec baseline.
Serving Efficiency: The unified approach reduced end-to-end serving latency by 11.1% and increased Queries Per Second (QPS) by 63.6% compared to a naive separate deployment. The KV-cache reuse alone provided a ~2.4x speedup for the ranking pass.
Online Engagement: In A/B tests on "Board More Ideas" and "Notifications":
- Board More Ideas: Achieved a +0.95% lift in surface saves and a +0.08% lift in site-wide saves.
- Notifications: Achieved a +0.91% lift in push opens, a +1.72% lift for dormant users, and a +0.09% lift in Weekly Active Users (WAU).

Significance
The paper claims that UniPinRec demonstrates that the "harder problem" of unifying input formats, training, and serving infrastructure is solvable in production. By moving beyond architectural unification to full-stack unification, the system eliminates the redundancy of encoding user history twice, reduces costs, and improves model quality through signal sharing. The authors position this as a practical, incremental path for industrial systems to adopt generative retrieval and ranking without sacrificing operational flexibility or requiring a complete overhaul of the recommendation funnel. Future work includes extending this unification to the L2 ranking stage and other surfaces like Search and Ads.