GENERATIVE · 2026-10-27
Generative — Paper Daily, 2026-10-27
The generative track has 8 papers today, 4 with deep dives. Pinterest's PinRec serves the home feed, search and related items with one generative retrieval model, and the key is generating toward the target action. OMG finds that generative recommenders do poorly when few history items share the target's category, and inserts a small iterating network between encoder and decoder that turns behavior-leaning representations into what code generation needs. CoFiRec codes Semantic IDs coarse to fine by category, text and collaborative signal. Another paper distills a sequential model's item distribution into a parallel generative model. Teams doing Semantic ID retrieval can start with PinRec and OMG.
4.45
LEAD STORY ·
PinterestRecommender systems usually run a separate retrieval model for the home feed, search and related items, because each surface wants something different: saves on the home feed, outbound clicks to buy on search, and continued browsing on related items. This Pinterest KDD 2026 paper replaces all three with one generative retrieval model, PinRec. It is pretrained on site-wide behavior, fine-tuned on each surface's logs, and told at generation time which action to aim for. Conditioning on the target action matters far more than unifying the model. Removing that condition cuts offline recall on the three surfaces by 7.5% to 12.6%. Added to search as one retrieval source online, it raised saves by 3.88%.
4.35
Generative recommenders that write the next item's Semantic ID code by code have a systematic bias. This KDD 2026 paper from South China University of Technology finds that the fewer items in a user's history share the target's category, the worse they do against classic behavior-only sequential models; in the worst group, TIGER's NDCG@10 trails SASRec by 30%. The cause: the encoder learns which item comes next, while the decoder must write its code. OMG inserts a two-layer network between them; each step adds the previous state to the encoder's final state and runs it through the network again. Each step predicts both the next item and the first code, with the latter weighted more each step, and all intermediate states go to the decoder. That one part lifts metrics by 14.5% on average while adding only 0.16% compute.
4.35
When shopping online, people usually look at a broad category before picking a specific item. Yet most generative recommenders build Semantic IDs by mashing an item's category, title, description and collaborative signal into one vector before quantizing it, so that coarse-to-fine structure is lost. This UIUC and Meta paper proposes CoFiRec, which splits an item into category, text and collaborative levels, codes each separately, and generates them coarse to fine. Across several public datasets and backbones, it beats the strongest traditional and generative baselines by 7.27% on average in Recall and NDCG. On cold items the gain reaches 10.0%, versus 4.7% on warm items.
4.25
In generative recommendation with Semantic IDs, training only tells the model which item came next, so it learns nothing about how likely the other items were given the same history. A classic sequential recommender can provide that distribution, but it scores items while the generative model predicts codes position by position, so the knowledge is hard to distill directly. This paper from HKUST and others proposes two distillation objectives for parallel models that predict all code positions at once. One rebuilds a catalogue-wide item distribution from the student's code predictions and aligns it with the teacher; the other spreads the teacher's item probabilities onto each code position. On average, the first lifts Recall@10 by 15.39% and the second lifts NDCG@10 by 8.28%.
4.25
Generative recommenders retrieve by writing an item's Semantic ID code by code, and every beam-search step runs through the large backbone, so the cost grows with more candidates and longer codes. It is often the main expense. This paper proposes Deco, which separates reading the history from writing the code. The backbone reads the history once and condenses it into a vector for the next item, and a light cached decoder runs the whole beam search, with codes and the prefix tree unchanged. On seven e-commerce, movie and music datasets, accuracy matches or beats the original model. Under a 50 ms latency budget, throughput is 3.2 to 8.6 times higher.
4.05
Generative recommenders are often trained with supervision first, then aligned to user feedback on whole slates with reinforcement learning. To keep training stable, each update may change the probability of generating a slate only by a fixed ratio set per sample, which smothers rare positive feedback on cold items and barely punishes repetitive negative slates. This Xiaohongshu paper proposes SAGE, which only changes the optimizer: positive slates get a higher cap on how far their probability can rise. Repetitiveness is measured by the entropy of the slate's item categories, and the further a negative slate's entropy falls below its recent average, the further its probability may be pushed down. On Amazon data, top-K accuracy holds while offline cold-start recall rises by about 90% to 101% and diversity by about 11%.
3.95
TencentLetting LLM agents rewrite code to search for better recommenders runs into two problems in generative recommendation. Retraining the tokenization, pretraining and alignment stages costs very different amounts, yet agents spend budget as if it were uniform. And a good upstream change often makes metrics worse until downstream stages adapt, so it gets discarded too early. This Tencent paper proposes TGR-EvoDAG, which keeps several improvement branches alive in a directed acyclic graph. Each node is an edited version of the code that records which pipeline stages the edit touched and its retraining GPU hours. Upstream edits that first hurt metrics are kept until downstream edits adapt, then the branches are merged, and budget goes by expected gain versus retraining cost. On Tencent's commercial media platform, the evolved model raised online click-through rate by a statistically significant 0.40%.
3.91
The Semantic IDs common in generative recommendation are opaque strings of codes: it is hard to see what the model generated or why, and the LLM's own reasoning goes unused. GUIDER splits recommendation into two steps. The LLM first reasons in text about which interest circles the user may want next. Circles are interest topics described in words and must come from a fixed set; items are then matched within the chosen topics. In reinforcement learning, invented circles outside the set lose reward, and so does overlong thinking. Large-scale online A/B tests on Kuaishou show steady gains in average app time and more exploration beyond users' existing interests.
KDD 2026
RecSys 2026