GENERATIVE · 2026-10-24
Generative — Paper Daily, 2026-10-24
All seven papers today deal with practical questions in generative recommendation: how to build semantic IDs, how to generate a whole result list, and how to cut reasoning cost. UniRec and CSG4Rec change how semantic IDs are constructed—the first keeps a few popular codes from taking most exposure, the second encodes collaborative signals into the IDs; Kuaishou's GLASS feeds long-term behaviour into generation; Netflix's GenPage generates homepage rows and titles one by one with a single decoder; Kuaishou's OneLatent compresses the "write reasoning first, then output the SID" step of an LLM recommendation backbone into 3 latent tokens, reaching 17.49 times the throughput of explicit reasoning. The other two are LLM-based conversational recommenders: one proposes new evaluation metrics, the other lets the model decide each turn whether to ask a question or recommend.
4.65
The Shopee authors measured three-level semantic IDs (SIDs) built by standard RQ-KMeans: the top 10% most-exposed full codes took 89.62% of exposure, so a generative recommender's probability mass concentrates on a few popular codes. UniRec has the model first generate the item's level-2 and level-3 category tokens and only then the three SID levels, and when building SIDs it gives every code an exposure-weighted capacity limit. Once the category is generated, each later SID level only has to choose within that category. Offline HR@50 is 22.6% higher (relative) than OneRec-V2; online, against the old multi-stage pipeline, orders rose 4.76% and GMV 5.60%.
4.60
KuaishouWhen an LLM serves as a recommendation foundation model, generating a natural-language reasoning trace (CoT) before outputting the item's semantic ID (SID) makes the output long and cuts throughput. Kuaishou's OneLatent inserts 3 learnable special tokens between the user context and the target SID: training first makes the model reproduce a teacher LLM's reasoning text after these 3 tokens, then switches to predicting the SID directly after them; at serving time only these 3 tokens are fed and no reasoning text is generated. The teacher writes reasoning from three angles: inferring intent, scoring the clicked item, and explaining why it was clicked. On the industrial dataset, SID@64 is 17.44% higher than OneReason-CoT's reason-then-answer mode; but against the strongest baseline trained on the same teacher reasoning and answering directly, the gain is only 1.58%, so most of the improvement comes from the teacher reasoning data. The reasoning slots are fixed at 3, so decoding does not grow with reasoning length, and throughput is 17.49x that of explicit reasoning. A local-services advertising A/B test estimates a 9.6% revenue lift.
4.45
The Netflix homepage consists of rows, each with a theme and a set of titles, and it used to be assembled by several stages such as a row model and an entity model. GenPage encodes the user's viewing history, device, time of day and so on as tokens forming a prompt, and a ~200M-parameter decoder-only Transformer generates row and title tokens one by one, left to right and top to bottom. Every row and every title is a single token, so business rules can be written directly as a candidate mask at each step. In an online A/B test the core metric used for launch decisions rose 0.24% and end-to-end latency fell 20%; offline, enriching the prompt reduced the loss more than scaling the model from 120M to 900M parameters.
4.20
Because attention cost grows with the square of sequence length, generative recommenders usually consume only short behavior sequences. GLASS, from Kuaishou and Tsinghua, uses the long-term history in two places: before generating the first SID level, it computes the cosine similarity between every long-term history item and the center of each first-level code, counts the items per similarity tier, and compresses this into one interest token fed to the encoder; after the first SID level is generated, it uses that code to pull long-term history items sharing the same first-level code and fuses them through a gate to help generate the remaining two levels. The first-level code fixes the broad category, and the later two levels must tell items apart within it, so what they need is same-category history. In a two-week Kuaishou short-video A/B test, app usage time per user rose 0.052%.
3.95
The semantic IDs used in generative sequential recommendation are usually produced by quantizing item text embeddings layer by layer, so items with similar text get similar codes; collaborative information such as "often bought by the same users" never enters the codes. CSG4Rec adds a contrastive loss when training the tokenizer: the quantized vectors of two items with similar collaborative behavior are pulled together and those of dissimilar items are pushed apart, so that items with similar codes are similar in both text and user behavior. It also aggregates the embeddings of each user's historically interacted items and quantizes them into a user semantic ID, which is fed into the generative model together with the item semantic IDs. The authors report gains over existing baselines on three benchmark datasets.
3.85
In conversational recommendation, Recall@K and BLEU count an LLM's movie suggestion as wrong whenever it is not the exact labeled answer, even if both are action comedies. The authors find that the labeled answer appears in the top 10 recommendations only 39% of the time, while another 53% of recommendations share at least a genre with the answer. They build one unified pipeline: a prompt extracts titles, genres and actors from the user's words, these are used to filter candidate movies from a knowledge graph, and an LLM generates the reply; six instruction-tuned models are compared on two movie dialogue datasets, ReDial and INSPIRED. Two new metrics, GHR@10 and SoftGHR@10, score whether the top 10 recommendations overlap the answer in genre, with SoftGHR also crediting partial overlap. The authors conclude that recommendation quality and fabrication depend on the knowledge graph and dialogue context, not on model size.
3.85
In conversational recommendation, users often cannot say what they want, and asking about attributes turn after turn drags the dialogue out. DESIRE trains no model; it uses LLMs such as GPT-3.5-Turbo with prompts to handle each turn: it first summarizes the dialogue context, the user's historical preferences, attributes retrieved from the item database and this turn's accept or reject feedback into a needs description, then decides whether to keep asking or to recommend; when asking it picks one of five question types, when recommending it picks one of six persuasion strategies such as presenting evidence or social proof, and item attributes in the reply may only come from retrieval results. Evaluation uses an LLM-simulated user. On LastFM the success rate within 15 turns (SR@15) is 0.583, versus 0.544 for the multi-agent baseline MACRS; removing the needs-summary step drops it to 0.327, showing that the success rate mainly comes from summarizing multiple signals first and then deciding whether to ask or recommend.
RecSys 2026
Netflix