GENERATIVE · 2026-10-19
Generative — Paper Daily, 2026-10-19
Community search over short UGC posts, questions, answers, and comments; prioritize relevance (nDCG), CTR, semantic matching, query understanding, retrieval, ranking, and practical LLM use.
4.45
LEAD STORY ·
YandexYandex Gryphon-v2 encodes user history once, generates semantic item IDs, and ranks the resolved items with a small shared-state head. A larger teacher supplies training targets but stays out of serving. Replacing the Yandex Music cascade increased active users by 1.41% in an online experiment.
Key insight: Distill on candidates the generator actually produces as well as logged impressions. This teaches the ranking head its serving distribution while retaining exposure-based supervision; sharing the history encoder then removes redundant online work.
3.95
Vela turns real behavior histories into tasks and accepted training trajectories rather than treating logs only as next-click labels. Fine-tuning on this supervision improves user-related tasks by up to 11.5 points over the base model.
Key insight: Behavior logs can supervise preference and intent inference, not just item prediction. Task filtering and trajectory acceptance matter because an observed action alone does not provide a reliable reasoning trace.
3.85
GoogleUniLang adds grounded machine-native symbols to a pretrained LLM vocabulary so text and item codes share one autoregressive objective. On MovieLens-20M, its reported NDCG@5 improvement over the strongest tabulated baseline is about 151.7%.
Key insight: Make structured symbols first-class tokens instead of verbalizing every identifier. Grounded initialization matters, and gains are metric-dependent: the Beauty Recall@10 result does not beat the strongest baseline.
3.80
SnapThis DPA evaluation method compresses behavior into intent representatives and uses an LLM judge, later distilled into a student, to score user-product relevance. One evaluated content retrieval strategy achieved a 4.98% online CVR increase; that is a strategy result, not a causal gain from the judge.
Key insight: Preserve multiple intents when compressing user history. Evaluate whether an offline judge tracks business outcomes, while keeping the effects of the evaluated retrieval strategy separate from those of the evaluation tool.
3.80
XXJTLU
University of Liverpool
Alibaba PsyBer-Agent simulates evolving user states and calibrates its trajectories against interaction logs. A lightweight policy handles most decisions, with occasional LLM fallback. Its five-million queued-task test reports a 95.18% success rate.
Key insight: Separate state dynamics from expensive language generation. The queued-task count is a workload size, not concurrent users; behavioral fidelity, model-defined state consistency, and system throughput answer different questions.
3.70
A retail experiment tests combinations of relevant and irrelevant provider context. For one customer, held-out recommendation appropriateness falls from 0.94 with a smaller context set to 0.88 with all signals. Extra context can disrupt retrieval even when the model does not adopt it.
Key insight: Choose context by the current intent and enforce operational constraints throughout generation. The study uses two customer profiles; its contribution is a context-selection framework.
3.60
AURA supervises intermediate reasoning by how well embeddings separate relevant documents from hard negatives. A small utility estimator selects the stopping point. Under matched deployment conditions, adaptive stopping reduces latency from 4.6 to 0.76 seconds versus full-trajectory reasoning.
Key insight: Stop when additional reasoning no longer improves retrieval discrimination. Intermediate utility supervision gives a more direct control signal than rewarding only the final retrieved result. This is a retrieval-track supplement.
3.45
TIDE attributes delayed recommendation feedback to the memories used by a content-generation agent, then reinforces, modifies, or evicts those memories. In offline temporal replay for e-commerce marketing, it reports a 7.75-percentage-point Memory Evolution Gain.
Key insight: Evaluate memory by its incremental utility on strictly future tasks. Credit delayed outcomes to the memories actually used, while accounting for timing and context rather than reinforcing an experience after a lucky click.
3.45
BBUPT
Southeast University
Xiaohongshu LIFT factorizes an interaction into Request, Item, Context, and Action states in causal order. Retrieval and ranking read different states from one shared history model. Its Joint Score improves by 3.6% over the strongest evaluated baseline on Taobao.
Key insight: Share the history representation while preserving what each stage is allowed to know. State boundaries help prevent future-information leakage. Joint Score summarizes a retrieval-ranking trade-off; it is not an online CTR result.
3.35
Input-adaptive sparse supervision trains an LLM retriever to retain salient lexical activations without a fixed support budget. Compared with Lion-sp, the paper reports 6.4-fold smaller index storage and competitive retrieval effectiveness.
Key insight: Allocate lexical support by cumulative activation mass rather than assigning every input the same budget. Evaluate query computation and index search together when comparing end-to-end efficiency. This is a retrieval-track supplement.
EMNLP 2026