GENERATIVE · 2026-10-24
Generative — Paper Daily, 2026-10-24
All 7 papers today tackle problems in deploying generative recommendation: how to build semantic IDs, how to generate a whole page, and how to cut inference cost. UniRec and CSG4Rec improve how semantic IDs are built, the first keeping popular codes from taking all exposure and the second putting collaborative signals into the IDs. Kuaishou's GLASS brings long-term behavior into generation, Netflix's GenPage generates homepage rows and titles one by one with a single model, and Kuaishou's OneLatent compresses reasoning before the SID into a few latent tokens. Two more papers study LLM-based conversational recommendation, one improving evaluation metrics and one letting the model decide each turn whether to ask or recommend.
4.65
Generative recommendation generates a product's semantic IDs layer by layer, and before the first layer it has seen no product-side information at all; this is its gap with discriminative ranking models. When a discriminative model scores a candidate it sees the product's category, shop and brand directly and crosses them with user features. UniRec has the model first generate the product's category and then generate the semantic IDs conditioned on that category, which puts back into the decoding path the product features a discriminative model can use. Launched online, it raised orders by 4.76% and GMV by 5.60%.
4.60
KuaishouWhen a large language model serves as a recommendation foundation model and first writes a natural-language reasoning trace before outputting the next item's semantic ID, the reasoning text is far longer than the semantic ID and holds back both decoding time and throughput. OneLatent appends 3 fixed special tokens after the user context and trains in several steps. First the model learns to reproduce a teacher LLM's reasoning text after these tokens, then it is switched to outputting the target semantic ID directly after them. At serving time only these tokens are appended and no reasoning text is generated. On industrial data its semantic-ID hit rate within the top 64 candidates is 17.44% higher than reason-then-answer, and its inference throughput is 17.49 times that of explicit reasoning.
4.45
The Netflix homepage used to be assembled by a multi-stage recommender that retrieved candidate rows and titles, ranked them separately and then built the page; GenPage instead generates the whole page with one decoder-only Transformer. The homepage is made of rows, each row a theme. User history, account attributes and request information are encoded as a prompt, and the model generates row tokens and title tokens one by one, left to right and top to bottom. Training copies the large-language-model recipe: first pretrain with next-token prediction on homepages that received positive feedback online, then post-train with weighted binary classification or reinforcement learning. The former labels each shown row or title by whether user feedback was positive and weights it by the reward size; the latter optimizes the sum of rewards of all titles shown on the page. In a 14-day A/B test it raised the core user engagement metric Netflix uses for launch decisions by 0.24% and cut end-to-end latency by 20%.
4.20
Generative recommendation generates the next item's semantic IDs one code at a time, and attention cost grows quadratically with sequence length, so these models usually take only the most recent few dozen behaviors and leave the long-term history unused. GLASS uses the long-term history differently for each semantic ID level. Before generating the first-level code, it averages the item vectors of each first-level code into a prototype. It then splits the cosine-similarity range into equal bins, counts how many long-term history items fall into each bin for each prototype, and passes the concatenated counts of all first-level codes through a small network to get one interest token fed to the encoder. After the first-level code is generated, it uses that code as a retrieval key to pull items of the same broad category from the long-term history, which help generate the last two codes. On TAOBAO-MM, R@1 is 21.57% higher than the strongest baseline; in a two-week A/B test on a short-video platform, recall rose 3 percentage points in absolute terms.
3.95
Generative recommenders encode each product as a sequence of semantic IDs, and this work finds that learning behavioral similarity directly in the space where the IDs are built works better than aligning it afterwards, beating strong existing baselines on three common datasets. Existing methods first build the IDs from product content and then align the ID vectors with embeddings learned by collaborative filtering. The two kinds of vectors live in different spaces, so forcing them together works poorly and misses which products tend to be bought by the same users. Instead, collaborative filtering first finds, for each product, products with similar and dissimilar behavior, and then the ID vectors of the former are pulled closer while those of the latter are pushed apart. It also combines the text vectors of the products a user has interacted with into user semantic IDs, built the same way, and feeds them to the generative model together with the product IDs. Product IDs should reflect not only content but also who buys the products together.
3.85
When an LLM chats with a user and recommends movies, the usual Recall@K and BLEU only check whether a recommendation is the exact movie the user finally picked in the dataset, yet many recommendations that miss it still suit the user's taste. On ReDial and INSPIRED, two movie-recommendation dialogue datasets, this study found that the reference film appears in the top 10 recommendations only 39% of the time, while 53% of recommendations share at least a genre with it. So the study adds genre-based metrics that count a recommendation as correct when the top 10 include a film of the same genre as the reference, and then compares six instruction-tuned LLMs in the same pipeline. A bigger model does not necessarily recommend better; knowledge-graph access and how much context the model gets matter more. Evaluating conversational recommendation means checking whether recommendations fit the user's preferences, not only whether they hit the single reference answer.
3.85
In conversational recommendation, first inferring what the user really wants from a few scattered remarks, and only then deciding whether to ask or to recommend, makes success much more likely. On LastFM music recommendation, with GPT-3.5-Turbo talking to a simulated user, adding this step raises the share of sessions that succeed within 15 turns from 0.327 to 0.583. For example, the user says they want something calm, adds that they don't want a fast tempo, and has a history of jazz and ambient music. DESIRE has the LLM put these remarks, the listening history, retrieved song facts such as a jazz track having a slow tempo, and the songs just rejected together to infer the real need. The LLM then judges whether this is enough to recommend. If not, it asks, for example directly or by having the user compare two options; if so, it recommends and persuades with evidence or emotional appeals, and grounds the song attributes it mentions in the retrieved facts.
RecSys 2026
Netflix