GENERATIVE · 2026-11-02
Generative — Paper Daily, 2026-11-02
The generative recommendation track has 10 papers today, 4 with deep dives, and the common thread is running generative recommenders in production. Kuaishou's GR4AD-X prunes ad candidates by revenue rather than generation probability. Alibaba's PFA has all beams share one ranking of important behaviors when decoding long histories, which saves most of the attention compute. Taobao's UGR generates the final list directly from about a thousand cross-business candidates, skipping separate per-business ranking followed by blending. RoFiGR uses collaborative signals to keep the first SID codes pointed in the right direction. Other papers train a query generator from retrieval results, make SFT tree-shaped, federate recommender agents and build a generative user embedding at DoorDash, and one audit shows that LLM recommendations shift with candidate order. Teams working on ad retrieval can start with GR4AD-X and PFA Teams working on blended feeds can start with UGR.
4.60
LEAD STORY ·
KuaishouKuaishou Generative retrieval for ads has a basic mismatch: the model picks ads by generation probability, while the platform cares about revenue. Kuaishou measured it. The top ten ads by generation probability and the top ten by eCPM barely overlap, with a hit rate of only about 0.02. GR4AD-X changes the deployed GR4AD in three places. When building SIDs, the first level keeps the few nearest cluster centers, so an ad has several paths and is not lost when one prefix is pruned. Capacity comes not from a bigger backbone but from more input signals, a larger encoder, a larger embedding for the SID vocabulary and so on. Each SID prefix takes the mean revenue of the ads beneath it as its value, and a value head at each level is trained to predict it. Pruning first keeps twice as many candidates, then uses RRF to merge the probability ranking with the value ranking before cutting. The changes were stacked one at a time in online tests, and together they raise ad revenue by 3.07% over GR4AD.
4.60
Alibaba
Zhejiang UniversityWhen a generative recommender decodes with beam search, every beam at every step computes attention over the whole user history, so longer histories mean slower serving. The authors at Alibaba found that attention concentrates on a few past behaviors as decoding proceeds, and which behaviors matter stays largely the same from step to step. PFA therefore looks at the full history only at the start of decoding. After that, the beams' attention is pooled, weighted by each beam's generation probability, into one ranking of important behaviors shared by all beams, and those behaviors' KV cache is moved into contiguous memory. At each later step the number of kept behaviors shrinks by a fixed ratio, for example by half per step. The important behaviors can be found early, so there is no need to reselect them at every step. In an online A/B test on an e-commerce ads platform, the user history grew from 128 to 512 behaviors, revenue rose 3.76%, and response time stayed below that of feeding all 512 behaviors directly.
4.60
Alibaba Generative ranking in e-commerce recommendation has a mismatch: the list the model learns to generate is not the list users finally see. One screen mixes organic content, ads and items from other businesses, and the usual setup ranks each business separately before a blending stage stitches them into the final list, changing positions and context. Taobao's UGR hands the roughly one thousand cross-business candidates of a request to a single generative model, which produces the final list item by item. Training goes beyond imitating logged exposure sequences: samples are weighted by post-exposure clicks and purchases, and the model also learns which of two orderings is better. Ads keep the existing allocation and GSP pricing rules. In an online A/B test on Guess You Like, transaction volume rose 3.29% and GMV rose 2.10%.
4.25
A generative recommender produces an item's SID one code at a time, and the first codes set the broad direction. One wrong branch early on and no amount of later precision can recover the relevant items. This is sharper in multimodal recommendation, where SIDs are usually built from image and text semantics, so the first codes may not reflect which items the same users tend to buy. RoFiGR aligns the first two levels of code vectors with frozen collaborative embeddings when it builds the SIDs. During generation, a sequential recommender acts as a teacher and tells the student model, at each step, how much probability each branch that leads to a real item should get. The first levels get a fixed, larger distillation weight, and the rest goes to whichever levels the teacher favors the correct branch more than the student does. On three Amazon datasets, RoFiGR beats the compared baselines on all five ranking metrics. Averaged over the 15 dataset and metric pairs, it is about 9.4% above the strongest baseline.
4.20
One way to use an LLM for recommendation is to have it write a few search queries from the user's history and pass them to an existing retrieval system. The catch is that logs record what the user clicked, not what query should have been written. The common fix imitates the clicked item's text, but such a query may not actually retrieve that item. R2QRec checks queries with a fixed retriever instead. It first generates many candidate queries and keeps only those that really retrieve the clicked item as SFT targets. The model then writes a set of queries, each retrieves a list, RRF merges the lists into one recommendation list, and GRPO rewards the set by where the clicked item lands in it. Only the query generator is trained; the retriever and item index stay fixed. On MovieTV, removing the retrieval check drops NDCG@10 from 5.07% to 4.03%.
4.15
Put candidate products into a prompt and ask an LLM to recommend one, and the order of the candidates by itself changes the pick. The authors built a RAG recommendation testbed on a real product catalog, shuffled the candidate order repeatedly for each query, and tested 8 models. Statistical tests confirmed that the recommendation depends on position for every model. The direction differs by model: GPT-OSS favors items listed first, while LLaMA 4 Scout avoids the first slot. It gets worse with more candidates. With 20 candidates, merely shuffling the order makes 58.1% of DeepSeek R1 7B's recommendations differ from the product it most often gives for the same query. The bias comes from the model itself and does not go away with a different temperature, so teams using LLMs for recommendation should randomize candidate order and audit for this regularly.
4.15
In LLM-based generative recommendation, the input is a long, text-heavy history while the output is an SID of just a few tokens. Training spends a lot of compute reading a whole history and then usually learns only one target item from it. Merge and Branch works on both ends. First, each history item is squeezed into a handful of tokens. The item's per-level SID embeddings are first combined into one vector. A few learned query vectors are each added to it and attend over the item's SID and text tokens. What each query picks up is added back onto that vector to form one token, so the input gets shorter. It then has the compressed history predict several target items at once; the targets share the same history like branches of a tree and cannot see each other. At decoding time all beams also share one history KV cache. The history is computed once and every branch still sends gradients back. At a history length of 512, beam search uses 61.8% less memory and runs about a tenth faster.
3.85
LLM recommender agents face a choice: pooling every user's memories raises privacy concerns, while isolated agents cannot learn from each other. FedRecAgent lets each client-side agent upload only experiences written in natural language, with no interaction logs and no model parameter updates. Each experience has three parts: what situation the user was in, how the next choice went (staying with the current interest, returning to an earlier one, or trying something new), and why. Specific items and personal preferences are stripped out. The server merges experiences whose situation and choice change both match, and keeps apart those where similar situations led to different choices. Each client then uses its local memories of preferences and choice habits to pick the most useful shared experiences. Across six datasets, the best variant beats the strongest baseline on NDCG@10 by 1.54% to 9.19%.
3.80
Ads ranking on food-delivery and local-commerce marketplaces needs strong personalization, yet user histories are sparse and there are hundreds of thousands of stores. DoorDash's ranker used to rely on hand-built count features and order-free pooled history, with no learned representation of a user's sequence. PALATE trains a GPT-style model to predict which stores a user will engage with next from their behavior sequence, and takes the hidden state at the last position as the user embedding. It is refreshed daily and fed to the ranker of every ad surface as an ordinary feature. Because repeat orders dominate, training gives more weight to stores new to the user, pretrains on clicks and then fine-tunes on orders. Offline it is the largest single-feature AUC gain. Online it shipped inside bundled ranker retrains, and on search ads revenue rose 0.94% and conversion rate rose 1.50%.
3.55
When predicting a user's next point of interest (POI), asking an LLM to generate the POI ID directly is unreliable: POIs are numerous, scattered and long-tailed, and their names carry little meaning. ANCHOR changes the order. The LLM first predicts what category of place the user wants next, which is easier because categories are few and stable. Candidates are then gathered from three sources: places in that category, places the user has visited, and places from the categories the user favors over the long term, recently and at that time of day. Finally a reranker that looks at the user's past trajectories picks the specific place. Guessing the category first and then the place is steadier than guessing the place outright. On check-in datasets from New York, Tokyo and California, ANCHOR beats the strongest LLM baseline, Think2Go, on Acc@1 by 7.56%, 11.21% and 1.94%.
CIKM 2026 Workshop
DoorDash
ICDE 2027