GENERATIVE · 2026-11-06
Generative — Paper Daily, 2026-11-06
Today’s generative recommendation papers separate training objectives from serving computation. BRIDGE restores item relationships behind SID identifiers, while LatentRec retains reasoning-training gains in a deployment mode that generates no reasoning. Other papers examine whether rewards concentrate recommendations on a few items, and whether factually grounded shopping answers yield workable plans.
4.55
LEAD STORY ·
Kuaishou
Generative retrieval predicts the next item's Semantic ID (SID) one code at a time from user history, but ordinary next-code training checks only the probability assigned to the correct code. When that probability stays the same, redistributing the remaining probability among incorrect candidates does not change this cross-entropy loss.
4.45
Generating reasoning before recommending items makes every request wait longer. Starting from user history during training, LatentRec repeatedly weights trainable token embeddings by similarity to the hidden state, averages them and feeds the result back into the same LLM. Each step predicts the target item’s first SID token, and the loss updates LoRA adapters; deployment removes the rollout module. On Beauty, compared with a fine-tuned textual-reasoning baseline, single-request latency fell from 10.80 seconds to 0.59 seconds and hit rate improved by a relative 5.0%. Multi-step supervision remains in the LoRA adapters, while serving generates items directly.
4.40
Learning an independent embedding for every item limits information sharing for sparse items; assigning items permanently to SID groups loses differences within each group. CAP-Mix gives each item a learned weighted mixture of shared representative vectors plus its own vector. After training, these can be materialised as an item table, so serving need not repeat the mixture computation. With the same SIDs and backbone, HR@5 improved by a relative 4.45% to 14.90% over a hard-assignment table. The practical idea is to learn how much items share, instead of choosing only between fully independent vectors and fixed groups.
4.30
Using a collaborative-filtering model to reward generative recommendations can concentrate first-place results on the same few items. In the Beauty SID experiment, the number of distinct items recommended first across users fell from 1317 after SFT to 837. The paper subtracts each item’s average reward across users, retaining its deviation for the current user; the count recovered to 1830, but hit rate also declined. Part of the reward reflects an item’s broadly high scores rather than personalised matching. Separating these components changes recommendation concentration, without guaranteeing that accuracy comes at no cost.
4.30
MetaA recommendation slate needs to account for items already selected, but rerunning the full model for every choice is expensive. MIS encodes candidates and display positions once, then feeds candidate, display-position and ordered selected-item representations into a lightweight head to choose the next item. Selection remains sequential without repeating full encoding. In a randomised online experiment, ranking-service CPU usage fell by 9.73% and watch time rose by 0.058%, while prediction-service CPU increased by 1.39%. Computation shifts toward reusable representations and cheaper sequential scoring, but percentages from the two services cannot be subtracted to establish total savings.
4.00
A more specific shopping query with an image can improve compliance with the current request while reducing the influence of a user’s established preferences. PMRecBench evaluates these separately: the target meets the query and has a high user rating, while competitors meet the query but have low ratings or no past interaction; the test checks whether the target ranks above them. In Gemma’s multi-agent experiment, adding an image-text query compared with history-only input reduced this win share by 4.6 percentage points on clothing. Better overall ranking can therefore conceal weaker use of preference history. Adding multimodal input does not by itself establish better personalisation.
4.00
A shopping assistant can understand requirements and state product facts correctly yet still propose an unworkable purchase plan. CARTBench fixes product evidence and evaluates replies along timestamped dialogue. Besides requirements and evidence, case-specific rubrics check whether advice supports a workable purchase decision. In the travel example, buying power adapters is deferred until the destination and charger specifications are checked. A rule violation or a required omission is a failure. Among replies passing all resolved need and evidence checks, 13.0% still failed a plan check. Factually accurate product claims therefore do not guarantee a workable recommendation.
3.95
Local recommendations must generate items available for the particular request. Filtering only after generation can spend computation on items the user cannot buy. This paper compiles the request’s eligible items into a SID-prefix index and keeps only branches that can still reach an eligible item during decoding. CPU index preparation overlaps prefill. A request uses one fixed eligibility snapshot rather than continuously tracking inventory changes during decoding. The contribution is to enforce availability inside generation for location-dependent supply, while still requiring the request set and item identifiers to remain consistent.
3.90
HHong Kong Polytechnic University
NTU A fluent reasoning trace is not necessarily a useful training example for generative recommendation. MUSE compares how each trace changes the target item’s full-catalogue rank relative to using no reasoning. Training imitates the most helpful trace and requires its generation likelihood to exceed that of the least helpful trace. With the same model on VideoGames, HR@10 improved from 0.0844 with randomly selected traces to 0.1020. Target items are used to select training examples; serving does not ask a teacher to judge each trace again. Recommendation ranking supplies a more direct criterion for choosing reasoning data than the fluency of the explanation.
3.60
When a sequential recommender repeatedly applies the same computation block, an intermediate step can recommend better than the final one. During training, DESRec selects the intermediate state with highest cosine similarity to the known target item embedding, then teaches the final state its score distribution over items. Serving still runs a fixed number of steps and returns the final state. In the CDs SASRec experiment, relative NDCG loss from the best step to the last fell from 31.79% to 21.52%. This reduces deterioration in later steps without giving the deployed model advance knowledge of the best step or eliminating later computation.