GENERATIVE · 2026-10-21
Generative — Paper Daily, 2026-10-21
Community search over user-generated posts, questions, answers, and comments; relevance, nDCG, CTR, query understanding, and language-model applications.
4.90
LEAD STORY ·
Kuaishou /
FudanKuaishouFudan Generative retrieval need not encode every available feature up front. AGR requests intent, location, or behavioral evidence while decoding item identifiers, then resumes generation with a compact observation. A Kuaishou online experiment reports a 0.205% increase in requests with strongly relevant top-four results.
Key insight: Feature value depends on the current identifier prefix. Learning when to request evidence can reduce redundant history encoding, although tool use still costs more than the feature-free baseline.
4.20
SSwinburne
Australian Catholic UniversityUUniversity of Tennessee
University of Tokyo
Deakin POI recommendation often lacks text or images. CATH selects complementary agents for each request, shares candidate-specific evidence, and adjusts a fixed base ranking. Across three naturally incomplete datasets, relative NDCG@20 gains over the strongest evaluated prior baseline range from 8.4% to 9.1%.
Key insight: Agent complementarity matters alongside agent count. Jointly optimizing group selection, ranking quality, and execution cost directs work toward candidates with missing evidence.
4.20
UUniversity of BariOOtto von Guericke UniversityLLeibniz GEI
A language model can score choices directly rather than generate an explanation first. Decision4Rec combines graph-grounded candidate knowledge, agent evidence, and a base ranker. A trained decision model improves NDCG@10 by 25–36% over the best trained baseline in three sparse product domains.
Key insight: Much of the gain comes from knowledge about existing candidates. Holding the candidate pool constant helps separate knowledge enrichment from coverage and model-size effects.
4.05
KAIST
LGSearCl builds conversational recommendation data from search and click logs. A configurable state machine constrains the two dialogue models, while item knowledge graphs ground their responses. The authors generate 3,538 dialogues and assess quality with an LLM judge.
Key insight: Observed clicks ground the target; a state machine governs the interaction. This makes generation controllable, while judge-rated dialogue quality remains distinct from real user acceptance.
3.80
AAIPher
Tsinghua
Peking University Drug-target retrieval also suffers from unknown associations being mislabeled as negatives. GenFish generates shared protein identifiers conditioned on a molecule, avoiding explicit negative labels for every unmeasured pair. Its benchmark catalog contains 20,431 human proteins.
Key insight: The transferable idea is how retrieval handles incomplete supervision. Protein experiments offer a useful false-negative audit design, although their objects and metrics differ from community search.
3.70
IntuitMMichigan StateAn unobserved interaction is not proof of dislike. The authors extract behavioral rules, screen negative candidates using statistical evidence, and use an LLM to interpret their business meaning. In an industrial task with four negatives per positive, test PR-AUC improves by 12.5% over random selection.
Key insight: Evaluate negative-label validity separately from downstream training gains. The distinction is useful when unexposed items are routinely treated as negatives in sparse click logs.
3.45
The model can start answering before the user submits a request. Streaming generation drafts from partial input, then verifies and revises as more text arrives. On HumanEval, ReFusion-8B reaches a 3.68× median speedup in post-submission wall time with unchanged accuracy.
Key insight: Residual waiting time differs from total computation. Gains depend strongly on task structure and are much smaller for dialogue than for specification-first coding tasks.
3.00
A benchmark score can change with the token IDs used to read yes/no answers. This audit finds that common extraction rules miss leading-space forms emitted by the model. Correcting the readout shifts adversarial POPE F1 by 6.13 percentage points.
Key insight: The measurement interface is part of the experiment. Normalize answer forms and token mappings before attributing a score difference to a model or prompting method.
3.00
Writing new associations can damage a recurrent model’s existing memory. KronLSTM uses a more localized matrix update to reduce interference. Its hybrid model reaches 37.40% average accuracy on a six-task real-world recall suite.
Key insight: Control how widely each memory update changes stored associations. This is a potential building block for long histories, with evidence currently coming from language and contextual-retrieval tasks rather than recommendation.
3.00
Tsinghua Combining recurrence and attention does not guarantee useful specialization. The authors find that the mixer learning retrieval first suppresses the other’s signal and introduce staged training. In one Gated DeltaNet hybrid experiment, Recall rises from 17.33 to 32.05.
Key insight: Architectural complementarity does not ensure complementary learning. Freezing interventions and layer attribution test whether the intended long-range component actually learns its role.