GENERATIVE · 2026-10-28
Generative — Paper Daily, 2026-10-28
The generative track has 8 papers today, 5 with deep dives, and the common thread is making generative retrieval cheaper to serve. Pinterest's PinGER lets each history item cost one token at serving time. Taobao's SLIDE moves beam search offline so the model decodes greedily once online. ECNU's MuSID-KD aligns the student with the teacher at every SID level when distilling a small model. Tencent's LoopGR reuses one decoder in a loop instead of adding layers. Octen-WebGR scales generative retrieval to 200 million web pages and measures what it and dense retrieval are each good at. Teams putting generative retrieval into production can start with PinGER and SLIDE.
4.75
LEAD STORY ·
PinterestPinterest Generative retrieval encodes each item as a multi-level SID and generates it level by level, but on the input side every history item costs several tokens. During training, Pinterest's PinGER places one dense-vector token after each item's SID and requires that any token reading another item may see only that item's dense token, never its SID. At serving time all SIDs in the history can be dropped, each item costs 1 token, and the output is unchanged. It turns input compression into a structural constraint at training time. Offline it matches TIGER's 1,024-token recall with 170 tokens, and after launch on Related Pins saves rose 0.62%.
4.60
Generative retrieval usually produces candidates online with beam search, and the beams never see each other; writing a whole list in one go by greedy decoding is the most direct alternative but works much worse. Taobao's SLIDE (a training method that distills offline search results into lists) moves search offline: a teacher runs beam search for every training request, concatenates the retrieved items in score order into one long sequence, and a student learns it, decoding greedily once online. It trades online search cost for offline labeling cost. With a 7B teacher, the 0.5B student reaches HR@10 of 0.4578, beating the 0.4498 of beam search at the same size.
4.50
When a SID-based generative recommender is distilled into a small model, the student is usually asked only to match the teacher's distribution over which item to recommend in the end. ECNU's MuSID-KD (a multi-granularity distillation method that works level by level over the SID) finds this is not enough: a student aligned only on items disagrees with the teacher on the order of candidate groups at intermediate SID levels in 53.8% of comparisons on average. The teacher's knowledge lives not only in the last-level item probabilities but also in the ranking of prefixes at every level. On three Amazon datasets its NDCG@10 is 8.31% higher on average than basic distillation, with the online inference structure unchanged.
4.30
Making a generative recommender stronger usually means adding layers, which means adding parameters and memory. Tencent's LoopGR (a generative recommendation backbone that reuses one decoder in a loop) runs the same decoder several times. Between passes, the decoder's BOS hidden state (its current judgement of the target) is added to every history position and a tiny GRU rewrites the history along the item order. The decoder keeps revising its prediction, and the history changes along with it. At equal compute it beats an encoder-decoder backbone with about half the parameters; on industrial data its Recall@10 is 0.6458, above the production model BARGE's 0.6277.
4.10
Completing a product bundle or playlist is unordered generation of a set of items, so a diffusion language model with bidirectional attention suits it, and it can reveal several tokens per step. Each item is a multi-level SID whose first token decides the coarse category, but with confidence-ordered revealing the model postpones that token, so generated items look alike; revealing too many at once assembles SIDs that do not exist. DPE-Rec (a diffusion recommendation framework for bundle construction) generates the category token first during reinforcement-learning rollouts with a higher sampling temperature at that token, and trains denoising not on random masks but only on the intermediate states that slow one-token-per-step decoding actually visits. Decide which way to branch first, then fill in the details. On set-matching metrics such as F1, it beats existing generative recommenders by 63% on average.
4.00
Generative retrieval has stayed below ten million documents. The Octen team's Octen-WebGR (a model that does web navigational search with generative retrieval) pushes it to 201 million web pages: every page is named by two tokens, a 4B model generates them directly from the query, and a table lookup returns the URL. On real navigational traffic from the month after training, it ranks the target first 59.6% of the time, against 46.5% for a dense retriever trained on the same clicks and reranked by clicks. Writing click history into the weights teaches the model how popular each page is, so it can pick the page users want among several similar pages of the same site. On rarely clicked pages dense retrieval is still better, so the two are complementary.
3.75
Chain-of-thought LLM recommenders first write a user preference note and then rank, but the note is never checked and the ranker never consults behaviorally similar users. ReflectRec (an LLM recommendation method that verifies preference notes by a round trip) has the model guess which movies the user watched from the note alone, and uses the share of the real history it recovers as the note's trust score. Low-scoring notes are regenerated and the highest-scoring attempt is kept, while movies that similar users watched next are injected into the prompt as a collaborative hint. Whether a preference description can recover the history is the test of whether it can be trusted. On MovieLens-1M, Hit@10 is 4.46 points above the CoT4Rec baseline.
3.15
Vector quantization underlies image and audio generation and even recommendation SIDs, but the common VQ-VAE codebook often collapses, leaving most codewords unused. FlowVQ (a flow-matching-based vector quantization method) first cuts a fixed simple prior distribution into cells of equal probability, one initial codeword per cell. A lightweight flow model trained together with the VQ-VAE then maps the prior space one-to-one onto the data's latent space. The map preserves probability mass, so each cell still holds the same share of the data after transport, and the codewords are the centers of the transported cells. Even code usage becomes a property of the map itself. On material structure reconstruction it reaches a 90.81% match rate with the whole codebook in use.