4.85
LEAD STORY ·
AlibabaAlibaba In Taobao's Pailitao, a user may photograph a shirt and ask for shorts to match it, which calls for a complementary item rather than a duplicate. The usual pipeline classifies intent first and routes to separate retrieval channels, so one wrong step breaks everything after it. Alibaba's report lets a 2B vision-language model read the raw image and instruction and generate product Semantic IDs directly, with no intent labels. On matching requests the excellent rate rises from 44.2% for the multi-channel pipeline to 74.3%, and online GMV grew 8.94% at full rollout.
4.55
CIKM 2026
Meituan
USTC
RUC
PolyUAAcademy of CyberGenerative ranking models usually score every candidate in one forward pass. Refiner, from Meituan and the University of Science and Technology of China, adds a few more steps. Each candidate gets a reasoning token, and each step predicts only a correction added to the original score. The base model stays frozen; only LoRA and a new output layer are trained. On Steam it lifts MTGR's AUC from 0.7698 to 0.7872, and a 7-day online A/B behind production MTGR raised CTR by 0.75%. Because it learns only residuals, the extra steps refine the ranking instead of overturning it.
4.55
Alibaba
Zhejiang University When click and value teachers are distilled together into a generative recommender, equal per-token weights are unfair. On TencentGR the student's KL to the value teacher at the first token is 3.13 times its KL to the click teacher, and the harder teacher switches to the click teacher on the last two tokens. KL-MOPD from Zhejiang University and Alibaba weights each teacher per token by its pre-training distance, at no extra training cost. In a 10% traffic online A/B test, ad revenue rose 1.63% and GMV 2.20%.
4.45
Cross-entropy tells a generative recommender only which token is correct, with no signal about how other candidates compare. REGER, from the University of Electronic Science and Technology of China, splits a conventional sequential model's scores over all items into branch preferences along the SID prefix tree and distills them along the true path. Added alone to TIGER, it lifts Beauty Recall@10 from 0.0673 to 0.0794 with unchanged inference time; removing it from the full model cuts Recall@5 by 10.9% to 18.0%. Two smaller modules add semantic pretraining and attention heads that look at adjacent items.
4.45
Most generative recommenders give every item an SID of the same length, so popular items keep paying for fine-grained codes, while shorter codes squeeze long-tail semantics. JevRec borrows the natural-language pattern that frequent words are short: popular items get short SIDs and long-tail items long ones. A dedicated head predicts the length first and a separate head generates content conditioned on it, reducing gradient conflict between the two goals. On three public datasets it beats all compared methods on every metric and trains and infers faster than TIGER.
4.15
When a semantic-ID recommender predicts a token at one level, it attends to historical tokens at every level. LESSRec's inference-time analysis finds that history at the target token's own level is more informative and other levels can interfere. So it encodes history separately per level, attends only to the matching level during prediction, and adds an item-level semantic sequence, fusing both through separate cross-attention branches. LESSRec+ gives each level its own smaller encoder at the same total parameter count to cut encoder compute, and both work across three public datasets and three tokenizers.
4.05
Existing generative recommenders train a dedicated SID vocabulary fitted to language-model representations of item descriptions. This paper notes that pretrained sparse autoencoders already expose language-model features with human-readable annotations, so feature indices can serve directly as the SID vocabulary. The API framework keeps each item's selected feature indices and relative activation strengths, and trains an adapter that predicts both to score candidates. It needs no extra vocabulary learning and can explain a recommendation through the annotated features behind it.
4.05
When an LLM recommender is post-trained with GRPO, one sequence-level reward is shared by every token, and each signal comes from a single user. RPSD has a teacher with privileged information re-evaluate the student's own response, and uses their per-token confidence gap to turn the sequence reward into token-level credit. The privileged information comes from similar users: time-aware optimal transport compares interaction trajectories across users, and a diversity-aware step keeps complementary ones as evidence. It reports state-of-the-art results on several real-world datasets.
3.85
When users describe needs in free text, preferences often qualify one another, as in a seaside city that suits elderly parents but is not too far away. Existing methods treat the text as flat keywords or hand it to an LLM to generate queries freely. PDG-F first has an agent parse the request into a directed acyclic preference dependency graph, then retrieves candidates from the open web along graph paths and ranks them by topological path. Across three English and Chinese domains it beats retrieval agents and frontier LLMs, mainly because the reachable candidate pool gets much larger.
3.35
Judging recommendation explanations by hand is accurate but expensive, while a single LLM judge is prompt-sensitive and inconsistent. MAREval splits the job across agents: a planner assigns roles with debate-style prompting, a moderator keeps the discussion on track, an arbitrator aggregates multiple rounds, and Monte Carlo sampling adds robustness. On a public benchmark and an e-commerce dataset it agrees better with human judgments and varies much less across repeated runs; it passed a large human-annotation quality gate where earlier evaluators fell short, and an online A/B showed significant gains in engagement and revenue.