GENERATIVE · 2026-11-09
Generative — Paper Daily, 2026-11-09
Today's generative papers split between recommendation and search. On the recommendation side, SIDRec trains Semantic ID construction together with click prediction, GCB uses Semantic IDs to predict several future user behaviors at once, and RecGroupMemory updates group memory from the previous day's feedback while leaving the recorded observations untouched. On the search side, several papers ask whom generative engines favor. Big brands get an edge, and the gains from rewriting copy are unstable once pages go live on the web.
4.20
LEAD STORY ·
Alibaba
When a large model plans a recommendation workflow, it can be given memory of how a group of users has behaved. The difficulty is that new transaction feedback arrives the next day and the advice should change, while the record of the group definition, the observation window and the statistics must not be rewritten along the way. RecGroupMemory splits the memory in two layers.
4.10
When a large model generates recommendation lists directly, the way the fine-tuning data is preprocessed noticeably affects the results. The authors fine-tune Llama-3-8B-Instruct on MovieLens-1M and compare choices. Even how a movie title is written matters. Moving a trailing article back to the front, so that Matrix, The becomes The Matrix, gives an HR@10 of 10.67%. Keeping the dataset's original form gives 12.47%. Balancing the training data to even out results across groups also depends on the attribute. Balancing by age narrows the utility gap between age groups and leaves overall utility almost unchanged. Balancing by gender lowers overall utility, and the utility gap between men and women does not shrink. Preprocessing choices should be judged on utility, output behavior and demographic sensitivity together.
3.90
Most sequential recommenders predict only a user's next item, yet user behavior is a chain of actions in which later steps depend on earlier ones. Generative Chain of Behavior (GCB) predicts several future behaviors directly. It first encodes each item as a Semantic ID of four discrete codes. The first three come from RQ-VAE quantizing the item vector level by level, with each level's codebook initialized by k-means clustering to avoid collapse, and the fourth only separates items whose codes collide, so semantically similar items share similar prefixes. A Transformer then generates the Semantic IDs of the next few steps autoregressively, each step conditioned on the steps already generated. On the Amazon Beauty and Cell Phones datasets, GCB is best or near-best on most multi-step metrics, with the clearest advantage on Cell Phones. The further ahead the prediction, the more every model degrades.
3.75
A Semantic ID (SID) is usually computed offline by quantizing a product's content vector, and it stays fixed while the ranking model is trained. This causes two problems. First, two products with similar content may attract different users, and the SID carries no such collaborative information. Second, how products are split into SIDs has never been trained against the click-prediction objective. SIDRec initializes a collaborative representation from the embeddings of a trained ranking model and keeps updating it during training, then fuses it with the content representation. It also lets quantization receive gradients. The forward pass still uses the discrete SID, while the backward pass uses a smooth probability approximation, so the click-prediction gradient reaches the quantizer and updates its parameters, which changes each product's SID assignment. On three public datasets it beats every compared method on AUC and GAUC, by up to 5.15% relative to the strongest baseline. Removing both the fusion and the end-to-end optimization lowers AUC by 7.19%.
3.60
When generative search recommends products, do big brands get an automatic edge? The authors queried three live generative search engines. Among 1,176 brands, those in the top fifth by popularity were mentioned and cited 1.4 to 2.8 times as often as the average brand, and turning on web search did not reduce this. In a controlled experiment, four models ranked five product pages. One page kept its product information but had its brand name swapped for an invented one, and its rank fell by 0.44 positions on average. Rewriting that page's text with three common generative engine optimization methods moved its rank by only about 0.02 positions. For a new brand, rewriting copy helps little. What matters is the impression of the brand that the model already carries.
3.45
When a large model must find out a hidden answer by asking questions over several turns, for example guessing which film a user likes, it often does not pick the most informative question and does not keep track of what it has already ruled out. The authors first build a Bayesian teacher that keeps a probability estimate over the answer each turn and asks the question with the largest expected information gain. They then fine-tune Llama-3.1-8B-Instruct on dialogues generated by the teacher. The student imitates the teacher's questions and writes its probability estimate as a scratchpad before deciding the next step. On three puzzle tasks, the success rate on decoding circuits rises from 0.00 to 0.93, and the average number of attempts to guess a word falls from 32 to 3.4. A 70B language model as teacher does not match the Bayesian teacher, and the learned skill does not transfer to a task family the student has not seen.
3.40
Generative engine optimization (GEO) rewrites web copy so that LLM-based search ranks a page higher. Most earlier evaluations simply fed the text to a model, and nobody had confirmed whether the gains survive on real web pages. LIVE-GEO publishes the rewritten product pages as live webpages and lets three commercial generative engines open them and rank six versions of each of a batch of fictional brands, namely the original copy and five rewrites made with five existing GEO strategies. No strategy wins reliably. For the same brand, the three engines pick the same top product for only about half to two thirds of brands. ChatGPT is also sensitive to the order of the candidate links, and the same strategy gets a Top-1 rate of 73.8% under one fixed order and 93.8% under the opposite one.
4.80
Generative retrieval outputs a Semantic ID (SID), but several products can share one SID, and which of them are finally retrieved directly affects online gains. RedMall is a generative model for e-commerce recommendation and ad retrieval. Besides changes to the architecture and training objectives, such as treating items that were shown but not clicked as negatives, it handles this step explicitly.