PERSONALIZATION · 2026-10-20
Personalization — Paper Daily, 2026-10-20
Community search over short user-generated posts, questions, answers, and comments. Priorities: relevance (nDCG), CTR, semantic matching, query understanding, retrieval, ranking, and useful LLM integration.
4.35
LEAD STORY ·
University of Amsterdam /
Johns Hopkins UniversityEncode articles separately, then learn which ones matter. LSNR combines weighted sparse article vectors with a user-level semantic expansion, improving MIND NDCG@10 by about 11% over STAR.
Key insight: Article boundaries matter beyond context length. A lexical path preserves evidence from individual clicks, while a semantic path adds aggregate interests. Gains with a single history article show that truncation is only part of the problem.
4.35
MetaPredictable Scaling evaluates ranking-model changes with reduced-scale proxies and estimates their full-scale effects using historical learning curves. Training compute falls by 88% and 79% on two industrial ranking models.
Key insight: A useful proxy preserves which changes are better, not merely a small average prediction error. Scaling must be chosen for each architecture: reducing embedding width alone can scramble the ordering of candidate changes.
4.10
Meta Meta represents ads with a stable semantic vocabulary that combines content and engagement signals. Integrated with its ads models, tokenization is associated with a reported 0.35% conversion lift, measured internally in February 2026.
Key insight: Cluster purity, distribution balance, and codebook use offer early diagnostics for tokenizers. Across tokenizer versions these measures correlate with downstream CTR accuracy. The conversion result belongs to the integrated system, not an isolated public tokenizer experiment.
4.05
Microsoft
UIUCUserRepGen learns user descriptions from feedback on reconstructed conversations. Evaluation covers 700 conversations and examines transfer to unseen tasks and population diversity, tying profile quality to the behavior a profile induces.
Key insight: Evaluate a profile through the behavior it produces, rather than its descriptive richness. Compare reconstructed conversations with real ones and train the profile generator to preserve decision-relevant details instead of incidental biography.
3.80
Google
University of TorontoReRankID tests rerankers that see only compact semantic item IDs. Untuned models often copy input order. Ranking-reward training approaches strong text-reading baselines whose prompts are about 20 times longer.
Key insight: Ranking quality and faithful explanations are separate capabilities. Compact IDs can carry behavioral signal without supporting accurate item descriptions. Audit input-order sensitivity and factual claims alongside ranking metrics.
3.80
A shared memory-retention policy misses user-specific reuse patterns. This work combines an offline prior with online exploration and delayed feedback. On PerMemBench-S at budget 500, the micro-average coverage score reaches 84.48.
Key insight: Discarded facts usually produce no future label, so observed feedback is biased toward retained facts. A retention policy must handle both selective observation and delay rather than label every not-yet-used fact as useless.
3.80
University of Michigan
KitwareVISTA infers latent values from a small set of survey answers and steers a frozen language model. On the World Values Survey, it retains 94% of the full-history gain after asking only 6% of the questionnaire.
Key insight: Ask for the preference information that most reduces uncertainty. Linking a user representation to both model steering and adaptive questioning can reduce elicitation effort; the reported gain is survey-based rather than an online ranking result.
3.70
UIUC This router learns individual quality preferences and uses a constraint queue to trade inference cost against satisfaction. Its real-model proof of concept uses 100 queries, with satisfaction outcomes still generated by a simulator.
Key insight: Track accumulated satisfaction shortfalls rather than trusting each isolated quality estimate. The queue raises the price of low-quality choices when feedback worsens. Guarantees depend on the preference model; the study does not establish live-user satisfaction.
3.50
AArcade.dev
This deployed system retrieves tools before an agent calls them. On ToolRet, its top ten results contain every required tool 57% more often than BM25, targeting the missing-tool failures that break multi-step workflows.
Key insight: Measure complete-set coverage alongside individual relevance. A multi-intent search can fail despite one excellent result if another required information type is absent. This is a transferable retrieval metric, rather than direct evidence about personalized CTR.
3.40
Sakana AIQD-Rec searches over item–rationale pairs, retaining pairs that are novel or better than nearby archived alternatives. Experiments span four task families: acquisitions, customer targeting, papers, and books.
Key insight: Archive an item together with its rationale to preserve distinct reasons for considering it. Local comparisons encourage diversity, but a larger archive is not automatically more accurate: some book-recommendation settings show no quality gain.
SIGIR-AP 2026
Salesforce
UCLA