GENERATIVE · 2026-11-04
Generative — Paper Daily, 2026-11-04
Almost every generative recommendation paper today is built around SIDs. Once the user history goes in, the model generates the next item's code one position at a time. Each paper targets one weak point in that chain. HERO finds that distillation passes the teacher's preference for popular items on to the student, and fixes it on 1688 with two teachers. UAGR puts value signals such as shares and full views into both the codes and the beam search scoring, and is fully deployed at Kuaishou. Another paper finds that the good results beam search misses are all lost at the first step, so it moves width to that step. ACT goes further and gives every creator its own token, avoiding SID collisions altogether. The common lesson is that how SIDs are built and how beam search prunes matter as much as the model itself.
4.85
On-policy distillation (OPD), where the student generates on its own and a teacher gives a full probability distribution at each position, makes generative recommenders train faster, but the student drifts further and further toward popular items. The bias comes from two places: the teacher's probabilities already favor the SIDs of popular items, and the student rarely generates the SIDs of unpopular items, so those paths get almost no supervision. HERO trains a popular-item teacher and an unpopular-item teacher. Each training example goes to the teacher that matches the group of the item the user actually clicked, and the true SIDs of unpopular items are also sent to the unpopular-item teacher. Only the student is served. In a 15-day A/B test on the Guess What You Like module of the 1688 homepage, GMV rose 3.412%.
4.85
Generative recommenders build a video's SID from its content alone, so a video that is often shared and finished and one that is rarely finished get nearly the same code if their content is similar. Beam search then prunes only by generation probability, so high-value candidates can be cut early, and later reranking cannot bring them back. UAGR fixes both ends. A few binary bits in the ID of the last SID token are set aside for the video's value level, graded from post-exposure signals such as shares and completions, and the SID length stays the same. During decoding, each time a token is added, the model uses the current state to predict watch time, shares and similar outcomes. These are weighted into a value score that decides, together with generation probability, which prefixes get pruned. It is fully deployed at Kuaishou and raised video watch time by 0.206%. Value has to count at pruning time, not only in later reranking.
4.85
Recommendation retrieval is usually split by objective into several channels, such as collaborative retrieval, long-tail retrieval and interest exploration, each with its own data and model. A team at Alibaba's Taobao and Tmall Group finds that one large language model can handle all of them and still beat the dedicated model of each channel. The input states in one sentence what this retrieval is for, and the model directly generates a set of items from tens of millions of candidates, with latency kept within tens of milliseconds. Beyond the offline comparison, on an online advertising platform it raised the core metric, advertising revenue, by 3%. One model that switches retrieval objectives on instruction can save the cost of maintaining many separate channels.
4.75
MetaWhen recommending creators a user does not follow yet, the mainstream approach quantizes creators into SIDs and has a language model generate them, and this paper finds the bottleneck is not the language model but the SID layer. Several creators land on the same SID, and only one can be picked by popularity, so recall is capped and tilted toward creators who are already popular. The quantizer is frozen, so training cannot change the creator representations either. ACT simply gives every creator its own token, trains it end to end, and at inference takes the creators with the highest logits. In a controlled comparison where only the token layer changes, creator-level Recall@15 is 19.8 points higher, and the SID approach's per-user inference latency is about 2.7 times that of ACT. With hundreds of thousands of creators, one token each is more direct than SIDs.
4.60
Generative recommenders generate SIDs position by position with beam search, and the usual way to miss fewer good results is to widen every step at once. This paper checks six full test settings of TIGER one by one. Every result that the model's own scores would place in the top 10 but that gets pruned is lost at the very first step. So the better deal is to concentrate width on the first step and narrow the later ones. With this allocation, about half the prefix budget (the total number of prefixes kept across all steps) reproduces the top 10 of a beam search twice as wide, in the same order, with no retraining. Beam width should go to the step where results are most easily lost.
4.45
A user's full history and their last few actions describe the same person, but they give different signals. Putting everything into one sequence dilutes the recent interest, while two separate streams fail to line up with each other. MV-Thinker uses one shared set of learnable query vectors with two independent readers to read the full history and the recent history separately. Because both sides use the same queries, their outputs line up one to one while each keeps its own long-term or recent information, and only then does SID generation start. Across four public datasets and two SID construction methods, it beats both discriminative and generative baselines. This is an academic architecture improvement with no online validation yet.
4.45
Generative recommenders are trained with the same loss weight on every SID token, and this paper finds that weighting tokens by position and rarity reliably improves recommendations. The reason is that the first few positions of an SID decide the broad item group, the later ones only refine it, and some tokens are very rare. This paper adds two kinds of token weights. One looks at position: the more candidate items a position rules out once it is fixed, the larger its weight, so earlier positions weigh more. The other looks at frequency: the less often a token appears in the training data, the larger its weight. During training it optimizes these two weighted losses together with the ordinary likelihood loss. Compared with the base TIGER model, Hit@5 improves 6.11% on average, and both popular and long-tail items get better. It only changes loss weights, not the model architecture, so the cost of adopting it is low.
4.35
Generative recommenders produce results with beam search, and a pruned prefix can never come back, yet common self-distillation training ignores which prefixes are about to be pruned. The teacher in self-distillation is a frozen copy of the same model that also sees the target item's title during training, and it supervises every state the student generates equally. SA-OPSD distills only on the states the student's own beam search visits, and favors states whose scores sit closest to the beam's cutoff line, the ones that are close to being pruned. It does better on one industrial and two public datasets, trains at lower cost than full KL distillation, and leaves inference unchanged.
4.35
Generative recommenders are trained only on interaction data and make little use of the semantic knowledge already inside large language models. The beam search likelihood ranks token sequences, which is not the same as how relevant an item is to a user. HiGenTrans brings LLM knowledge in at three places. Before generation, it adds an LLM-computed user representation to the start of the sequence. In the middle layers, it pulls each layer of the generative model toward the LLM layer at the same relative depth. At the end, a small MoE scorer combines semantic relevance, generation confidence and candidate-set information to rerank the beam search output. All semantic features are computed only from the user's existing interactions and can be cached offline. On three public datasets, NDCG@10 improves by up to 33.7%, and on every dataset NDCG rises more than Recall, so the gain mostly comes from ranking the right items higher.
3.80
When a large language model makes recommendations, it often follows the popularity it learned in pre-training, recommending what is popular rather than what this user likes. This bias lives in the parameters and is hard to remove directly. NPRec leaves the parameters untouched. It first generates several candidate text guidelines from the user's interactions, each describing a possible preference. It then trains a small model that predicts whether a guideline applies to a given user and item. Each guideline gets two such probabilities, one with the real user and item, and one with their individual features erased so only item popularity remains. Only guidelines whose first probability is clearly higher are kept, since one that applies on popularity alone is really about following the crowd. At inference the guideline goes into the prompt, so the model recommends and explains by following it. With guidelines written by people, the top-1 hit rate on MovieLens beats basic prompting by 95.6%. How good the guideline is matters most.
CIKM 2026
Alibaba
CMU
NTU
Nanjing University